Author(s): Tjeerd Driessen, Bart-Jan Van Der Spek, Marius Sokolewicz
Linked Author(s): Tjeerd Driessen
Keywords: Rio Magdalena, multiple linear regression, probabilistic forecasting models, quantile regression, water control
Abstract: The Canal del Dique (CDD) is a man-made connection between the R�o Magdalena near Calamar and the Caribbean Sea near Cartagena in Colombia. The Environmental Restoration of Canal del Dique project was initiated after a large flooding event that occurred in December 2010. A solution is chosen in which the inlet of water from the R�o Magdalena is regulated by a water control structure near Calamar. This provides a controlled dynamics of water levels necessary for environmental purposes while maintaining the bi-annual peak flows in the Canal system. For operational control of this water control structure, a forecasted water level on the Rio Magdalena near Calamar is required. This paper describes the development of multi-linear regression (MLR) forecasting models that predict water levels for different time horizons using 10 water level stations upstream. Stepwise regression is used to systematically add and remove input variables (predictors) such that the total regression model improves. A challenge in the selection of input variables is the presence of multicollinearity which refers to the (mutual) correlation between input variables. Stepwise regression is performed using time series from 1990-2010 (calibration). As expected, the accuracy is decreasing for larger time lags, but is considered very acceptable given the time lag: RMSE (Root Mean Square Error) of 2. 0 cm and 10. 3 cm for respectively the 1-day and the 5-day forecast models. Validation is done using the time series from 1980-1990 and provides even lower RMSE values which has led to the conclusion that the forecasting power of the models is fairly good. Quantile regression is used for probabilistic forecasting and provides the uncertainty of the prediction. These quantitative uncertainty bands are useful for risk assessments and robust decision-making and provide, in combination with the MLR models, a useful tool for the operation of the Canal del Dique system
Year: 2017