Machine-learning methods for stream water temperature prediction

Abstract. Water temperature in rivers is a crucial environmental factor with the ability to alter hydro-ecological as well socio-economic conditions within catchment. The development of modelling concepts for predicting river water and will be essential effective integrated management adaptation strategies future global changes (e.g. climate change). This study tests performance six different machine-learning models: step-wise linear regression, random forest, eXtreme Gradient Boosting (XGBoost), feed-forward neural networks (FNNs), two types recurrent (RNNs). All models are applied using data inputs daily prediction 10 Austrian catchments ranging from 200 96 000 km2 exhibiting wide range physiographic characteristics. evaluated input sets include combinations means air temperature, runoff, precipitation radiation. Bayesian optimization optimize hyperparameters all models. To make results comparable previous studies, widely used benchmark additionally: regression air2stream. With mean root squared error (RMSE) 0.55 ?C, tested could significantly improve compared (1.55 ?C) air2stream (0.98 ?C). In general, show very similar models, median RMSE difference 0.08 ?C between From both FNNs XGBoost performed best 4 catchments. RNNs best-performing largest catchment, indicating that mainly perform when processes long-term dependencies important. Furthermore, was observed hyperparameter showing importance optimization. Especially FNN model showed an extremely large standard deviation 1.60 due chosen hyperparameters. evaluates variables, training characteristics stream prediction, acting basis regional multi-catchment preprocessing steps implemented open-source R package wateRtemp provide easy access these approaches facilitate further research.