Potential application of wavelet neural network ensemble to forecast streamflow for flood management

Streamflow forecasting, especially the long lead-time forecasting, is still a very challenging task in hydrologic modeling. This could be due to the fact that the forecast accuracy measured in terms of both the amplitude and phase or temporal errors and the forecast precision/reliability quantified in terms of the uncertainty significantly deteriorate with the increase of the lead-time. In the model performance evaluation, the conventional error metrics, which primarily quantify the amplitude error and do not explicitly account for the phase error, have been commonly adopted. For the long lead-time forecasting, the wavelet based neural network (WNN) among a variety of advanced soft computing methods has been shown to be promising in the literature. This paper presented and compared WNN and artificial neural network (ANN), both of which were combined with the ensemble method using block bootstrap sampling (BB), in terms of the forecast accuracy and precision at various lead-times on the Bow River, Alberta, Canada. Apart from conventional model performance metrics, a new index, called percent volumetric error, was proposed, especially for quantifying the phase error. The uncertainty metrics including percentage of coverage and average width were used to evaluate the precision of the modeling approaches. The results obtained demonstrate that the WNN-BB consistently outperforms the ANN-BB in both the categories of the forecast accuracy and precision, especially in the long lead-time forecasting. The findings strongly suggest that the WNN-BB is a robust modeling approach for streamflow forecasting and thus would aid in flood management.