20th webinar of the series entitled "Hydrological data assimilation using neural networks"
Speaker: Marie-Amélie Boucher, regular member, Université de Sherbrooke
Presentation : See the slideshow » (PDF, 4 309 Ko)
Paper: M.‐A. Boucher, J. Quilty, J. Adamowski. (2020). Data Assimilation for Streamflow Forecasting Using Extreme Learning Machines and Multilayer Perceptrons. Water resources Research, 56, https://doi.org/10.1029/2019WR026226
Summary: Although the last 25 years of research have often shown superior performance of neural networks compared to conventional hydrological models, Abrahart et al. pointed out in 2012 that neural networks are rarely used in operational hydrological forecasting. One possible explanation is the "black box" nature of neural networks, which makes it difficult to interpret a good or bad forecast. Using neural networks for data assimilation would combine the computational efficiency of neural networks and the explanatory power of conventional hydrological models. Therefore, the basic initial hypothesis that the speaker wanted to verify with this work is simply that "it is possible to use an ensemble of neural networks to assimilate data in a simple hydrological model". The speaker will show in this presentation that, so far, this hypothesis holds true, at least for the watersheds where it has been tested. This brand new data assimilation method has limitations and certainly needs further refinement, but I will show that it compares favorably with a well-known data assimilation method, namely a "vanilla flavoured" particle filter (without calibration).
Language: French, with slides in English
New series of weekly CentrEau webinars "all about water management". These short 20-minute presentations will be offered live every Thursday at 1 p.m. followed by a short discussion period. CentrEau professors and researchers, as well as guest speakers, will present their research work and its importance in a semi-popularized manner. Watch the previous webinars on our YouTube Channel!
