Bootstrapped artificial neural networks for synthetic flow generation with a small data sample

Hydrological calibration of mechanistic watershed simulation models often requires several years of continuous flow data. Unfortunately, historical flow information is highly limited for many ungauged or recently gauged watersheds. Synthetic flow generation methods could be used to extend the available flow records at data-limited watersheds and to create a statistically reasonable synthetic flow series as a target for hydrological calibration. This work compares the performance of a bootstrapped artificial neural network (BANN), a maintenance of variance extension (MOVE) and a modified drainage area ratio (DAR) for synthetic flow generation with a small data sample. The bootstrap method is used to estimate the generalization errors of neural networks with different structures and to construct the confidence intervals for each flow prediction. Comparisons are performed with respect to a case study at Buck Mountain Run, Albemarle, VA, USA, given various small data sets of flow observations and flow predictors spanning different time periods. The results show that BANN outperforms MOVE and DAR in the low and medium flow predictions and that the best performance can be achieved by replacing the highest 10% of BANN synthetic flows with corresponding DAR predictions. Even with a small data sample, the combined BANN and DAR model (BANN + DAR) can be used to create an accurate synthetic flow series.