• Samuel Westarp
4. term, Water and Environment, Master (Master Programme)
Climate change is expected to result in increased rainfall intensities in Danmark and therefor the risk of flooding of rivers causing damage to urban areas on their way will also increase. Modelling of the effects on long terms is important for the physical planing of the areas surrounding the rivers. But short term forecasts of the water level of the rivers are essential for early flood warning systems.
This project has focused on short term forecasting models up to 36
hour lead-time for river water level based on machine learning investigating wether such models could benefit from soil moisture measured in real time. The soil moisture in the vadose zone affect the path that precipiation takes to the rivers and therefor it is assumed that knowledge of the soil moisture will help models predict the water level respons to precipitation.
The two rivers Elling Å and Romdrup Å in Nothern Jutland have
been chosen as cases for the project mainly due to data availability as these rivers do not cause issues with flooding.
It is found that measured soil moisture significantly explain some
of the variantion seen in the water level respons to precipitation events measured at station Brinkhus N in a small creek part of Elling Å. Similar but weaker results are seen for Romdrup Å although a calculatede estimate of the soil moisture, the so called ’drought index’ published by DMI, give similar results. The results transfer to the forecasting models build using a neural network architechture called TemporalFusionTransformer (TFT), as models for Romdrup Å trained with the drought index do not improove when soil moisture data is included. For Brinkhus N we find slightly better performance with soil moisture in the model.
The TFT-architecture produce forecasting models performning at
NSE > 0.9 with a 24 hour lead-time over 7 to 9 month of test data. The models are based on weather data published in 10km grid format published free by the Danish Meteorologic Institute, who also plan on releasing forecast data later in 2022. Models trained on data from Elling Å and Romdrup Å are succesfully
transfered to four different rivers of varying size and geologi and geography of the watersheds. It results in a big performance gain with little training
data compared to models trained from scratch.
Publication date25 Jun 2022
External collaboratorWatsonC
No Name vbn@aub.aau.dk
Information group
ID: 472477128