Daily river water temperature forecast model with a k-nearest neighbour approach

Water temperature is a key abiotic variable that modulates both water chemistry and aquatic life in rivers and streams. For this reason, numerous water temperature models have been developed in recent years. In this paper, a k-nearest neighbour model (KNN) is proposed and validated to simulate and eventually produce a one-day forecast of mean water temperature on the Moisie River, a watercourse with an important salmon population in eastern Canada. Numerous KNN model configurations were compared by selecting different attributes and testing different weight combinations for neighbours. It was found that the best model uses attributes that include water temperature from the two previous days and an indicator of seasonality (day of the year) to select nearest neighbours. Three neighbours were used to calculate the estimated temperature, and the weighting combination that yielded the best results was an equal weight on all three nearest neighbours. This nonparametric model provided lower Root Mean Square Errors (RMSE = 1.57 degrees C), Higher Nash coefficient (NTD = 0.93) and lower Relative Bias (RB = - 1.5%) than a nonlinear regression model (RMSE = 2.45 degrees C, NTD = 0.83, RB = - 3%). The k-nearest neighbour model appears to be a promising tool to simulate of forecast water temperature where long time series are available. Copyright (c) 2011 John Wiley & Sons, Ltd.