Predicting water levels in the Rhine river using Temporal Convolutional Networks
Keywords
Loading...
Authors
Issue Date
2019-07-26
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
To ensure that the people in The Netherlands keep dry feet, Rijkswaterstaat (a Dutch gov-
ernmental institute) wants to know whether deep learning models can be a valuable addition
to current water level prediction models. For this, we set as a reference point the multilinear
regression model LobithW, which Rijkswaterstaat uses to predict water levels up to two days
ahead in Lobith. We developed LobithAI, using Temporal Convolutional Networks. Due to
missing data, it was impossible to have a direct comparison between the two models. In-
stead, we used two di erent evaluation strategies. When compared on data in the same time
range, LobithAI outperforms LobithW on both RMSE and precision. When compared with
LobithW directly after its validation, LobithAI scores higher on the RMSE, but LobithW
has higher precision for predictions one day ahead. Extensions to LobithAI are also possible,
with the best performing extension being the increase of the number of predictions per day.
We also inspected LobithAI using activation adjustment, showing that the model mostly uses
the closest and most recent measurements. While the comparison between the models was
incomplete, the results from the evaluations and the
exibility of LobithAI make it a valuable
addition to the models of Rijkswaterstaat.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen