Efficient modelling of water temperature patterns in river systems – benchmarking a set of machine learning approaches

Many approaches for modelling river water temperature are available, but not one exist that can be applied without restrictions. The applied method depends on data availability, dominant processes, scales and transferability. Process-based models are currently the best way to evaluate detailed management scenarios on reach scale and to understand underlying processes.  Due to limitations of data availability, however, more simplified approaches are frequently applied, where different meteorological or hydrological time series are statistically related to water temperature (or in the simplest case only using air temperature). Here, machine learning methods could help bridging a gap by allowing for more complex relationships without setting prior assumptions. They are thus integrating reasonable processes and dynamics within the catchment by learning from given data. However, up-to-date machine learning approaches have rarely been used in this field until now.

This contribution analyses a set of machine learning approaches for large-scale river temperature modelling. Deep learning methods, random forests and boosting methods are compared with the performance of commonly used simple and multiple regression models. These approaches are tested on 10 catchments with different characteristics, human impacts (e.g. hydropower, river regulation) and time series lengths (10 to 39 years). They are situated in the Austrian Alps or flatlands with areas ranging from 200 to 96.000 km². Observed data including daily means of river water temperature, air temperature, discharge, precipitation and global radiation are grouped to simple and advanced sets of input variables to analyse possible data dependencies. 

In summary, we compare up-to-date machine learning approaches for their applicability in river water temperature prediction. By implementing necessary data preprocessing steps and machine learning routines in a R package, we aim to make these findings easily accessible and reproducible for the community. This tool provides an attractive approach for large-scale river temperature modelling, where the requirements for using process-based models are not able to be met. Future applications can include e.g. short and long term forecasting of river water temperature to find management options for balancing environmental requirements.