In:
Österreichische Wasser- und Abfallwirtschaft, Springer Science and Business Media LLC, Vol. 73, No. 7-8 ( 2021-08), p. 308-328
Kurzfassung:
Stream temperature is an essential environmental factor that has the potential to change both hydro-ecological and socio-economic conditions in the vicinity of a river. In order to calculate stream temperatures as a basis for effective adaptation strategies regarding future changes (e.g. due to climate change), robust modelling concepts are needed. This study investigates 6 machine learning models: Stepwise Linear Regression, Random Forest, eXtreme Gradient Boosting, Feedforward Neural Networks and two types of Recurrent Neural Networks. The models were tested on 10 Austrian catchments with different physiographic characteristics and input data combinations. The hyperparameters of the applied models were optimised using Bayesian hyperparameter optimisation. To compare the results with other studies, the predictions of the 6 machine learning models were compared with the results of linear regression and the well-know air2stream-model. Both Feedforward Neural Networks and eXtreme Gradient Boosting showed the best prediction results in 4 out of 10 catchments. With an average root mean squared error of 0.55 °C, the tested models were able to predict stream water temperatures significantly better than linear regression (1.55 °C) and air2stream (0.98 °C). In general, the results of the 6 models showed comparable performances with only a median deviation of 0.08 °C between the individual models. In the largest catchment studied—the Danube upstream of Kienstock—Recurrent Neural Networks showed the highest model performance, indicating that they are best suited when processes with long-term dependencies are crucial in the catchment. The choice of hyperparameters strongly influenced the predictive ability of the models, highlighting the importance of hyperparameter optimisation. The results of this study summarise the importance of different input data, models and training characteristics for modelling stream water temperatures based on daily means and is the basis for future model developments regarding regional stream water temperature prediction. The tested models are available for the research community via the open source R package wateRtemp.
Materialart:
Online-Ressource
ISSN:
0945-358X
,
1613-7566
DOI:
10.1007/s00506-021-00771-3
Sprache:
Deutsch
Verlag:
Springer Science and Business Media LLC
Publikationsdatum:
2021
ZDB Id:
1186984-7
ZDB Id:
2383304-X
