Machine learning for discharge prediction in the Alps

In the context of the Alpine Drought Observatory (ADO) project, a database of discharge measurements with more than 1400 gauging stations on alpine rivers with, on average, 35 years of records was assembled. This wealth of information constitutes an ideal source for data-driven discharge modelling with Machine Learning (ML). Discharge forecasting is relevant for many sectors related to the water cycle, such as agriculture and energy production. Moreover, appropriate river low streamflow prediction can improve preparedness for drought-related risks.

This paper proposes comparing two ML algorithms for discharge prediction using meteorological reanalysis and modelled snow variables over the gauging stations' catchment area as predictors. The selected meteorological variables are total precipitation, temperature, and potential evapotranspiration. ERA5 reanalysis [1] bias-corrected with quantile mapping and down-scaled to a 5.5 km grid is the source. The last predictor is the snow water equivalent (SWE), obtained with an adaptation of the SNOWGRID model [2]. All the predictors have a daily temporal resolution.

First, we build on existing work [3] with Support Vector Regression (SVR). The experiments aim at predicting the monthly discharge mean in the present and up to several months of advance. We evaluate the performances of the different approaches, investigate each input variable's importance for several test catchments with different hydrological regimes, and carry out trials with different temporal and spatial aggregations to find the best configuration.

We evaluate the prediction with the r² metric. Depending on the size and water management in the studied basin, results range from 0,7 to 0,85 for the present. We also perform the analysis based on discharge anomalies (computed as the deviation from the average discharge for the specific day) to erase the climatology effect. In this case, the r² metric ranges from 0,5 up to 0,7. For predictions of the future discharge, the model's performance decreases in about one month to the level of climatology. The SWE is a relevant predictor since the performance decrease is slower for larger basins with a nivo-glacial regime.

The results show the suitability of ML for discharge prediction on different kinds of alpine basins with up to one month of advance. The subsequent development will be to conduct a similar analysis with convolutional neural networks (CNN). This class of deep networks should allow the model to learn the spatial pattern in the input data.

 

[1] Copernicus Climate Change Service (C3S) (2017): ERA5: Fifth generation of ECMWF atmospheric reanalyses of the global climate. Copernicus Climate Change Service Climate Data Store (CDS), date of access. https://cds.climate.copernicus.eu/cdsapp#!/home

[2]: Olefs, M.; Koch, R.; Schöner, W.; Marke, T. Changes in Snow Depth, Snow Cover Duration, and Potential Snowmaking Conditions in Austria, 1961–2020—A Model Based Approach. Atmosphere 2020, 11, 1330. https://doi.org/10.3390/atmos11121330

[3]: De Gregorio, L., Callegari, M., Mazzoli, P. et al. Operational River Discharge Forecasting with Support Vector Regression Technique Applied to Alpine Catchments: Results, Advantages, Limits and Lesson Learned. Water Resour Manage 32, 229–242 (2018). https://doi.org/10.1007/s11269-017-1806-3