Functional hourly forecasting of water temperature

The paper describes the problem of forecasting water temperatures on an hourly basis using previous water and air temperatures as predictors. Both time series are decomposed using functional principal components, leading to low dimensional vector autoregressive modeling. The principal component scores mirror serial correlation, which is also incorporated in the model. The modeling exercise is motivated by and demonstrated with data collected in the German river Wupper, and the approach is contrasted to alternative routines which have been suggested in statistics and hydrology.