Predicting Daily Streamflow Using L-Moments and Neural Networks

Author(s): Worland, S.; Knight, R.; Asquith, W.

Various statistical methods have been evaluated and applied to estimate daily mean streamflow and other streamflow statistics for ungaged streams and regional characterization. We elected to estimate daily streamflow in ungaged basins using flow duration curves (FDCs), L-moments, and machine-learning. The method is an extension of the Q1P1P2Q2 method that uses streamflow at gaged locations (Q1, streamflow at the gaged site) to calculate a time series of exceedance probabilities (P1, exceedance probabilities at gaged site) that are used streamflow at the ungaged site) at ungaged locations. The workflow requires estimating a FDC for an ungaged basin—a step achieved using L-moments to summarize the distributional geometry of FDCs and statistical regionalization models. We regionalize the first four L-moments computed from 10-year blocks of daily streamflow data from 1950–2010 for 1,030 gaged-basins that span from southern Texas to Florida. The decadal approach results in 3,027 L-moment ensembles available for regionalization. Out-of-sample predictions are used to simulate method performance at ungaged locations. The specific steps are (1) calculate decadal L-moments at gaged locations, (2) use multi-output neural networks and 34 basin descriptors to regionalize L-moments to ungaged catchments, (3) parameterize an analytical flow duration curve at the ungaged locations using the regionalized L-moments, (4) select donor sites using distance matrices in basin-descriptor space, and finally, (5) use the donated probabilities (P1P2) to generate daily streamflow values at ungaged locations. Additionally, compensation for no-flow conditions is made through logistic-regression like modeling. Uncertainty is incorporated into the predictions using stochastic neural-network dropout to approximate a posterior distribution of L-moments and streamflow estimates.

Go back


Past Conference Archive