Abstract: In this work, we develop and test two artificial neural networks (ANNs) to forecast streamflow in ungauged basins. The model inputs include time-lagged records of precipitation and temperature. In addition, recurrent feedback loops allow the ANN streamflow estimates to be used as model inputs. Publically available climate and US Geological Survey streamflow records from sub-basins in Northern Vermont are used to train and test the methods. Time-series analysis of the climate-flow data provides a transferable and systematic methodology to determine the appropriate number of time-lagged input data. To predict streamflow in an ungauged basin, the recurrent ANNs are trained on climate-flow data from one basin and used to forecast streamflow in a nearby basin with different (more representative) climate inputs. One of the key results of this work, and the reason why time-lagged predictions of steamflow improve forecasts, is these recurrent flow predictions are being driven by time-lagged locally-measured climate data. The successful demonstration of these flow prediction methods with publicly available USGS flow and NCDC climate datasets shows that the ANNs, trained on a climate-discharge record from one basin, prove capable of predicting streamflow in a nearby basin as accurately as in the basin on which they were trained. This suggests that the proposed methods are widely applicable, at least in the humid, temperate climate zones shown in this work. A scaling ratio, based on a relationship between bankfull discharge and basin drainage area, accounts for the change in drainage area from one basin to another. Hourly streamflow predictions were superior to those using daily data for the small streams tested due the loss of critical lag times through upscaling. The ANNs selected in this work always converge, avoid stochastic training algorithms, and are applicable in small ungauged basins.
Abstract: Sustainable water resources management is critical in both developing and established communities; particularly with the challenges associated with surface and groundwater contamination and potential precipitation shifts resulting from climate change. We test and develop a method for forecasting short-term (daily) streamflow using data-driven Artificial Neural Networks (ANNs) to efficiently manage water resources during times of shortage and provide improved flood mitigation strategies. In this work, a generalized regression neural network is combined with a counterpropagation network (GRNN and CPN respectively). This hybrid network and its individual constituents are compared with traditional temporal forecasting methods (e.g. multiple linear regression and auto-regressive moving averaging or ARMA). Model inputs consist of antecedent precipitation, temperature and discharge, while the output is river discharge in space through time. A hierarchy of ANNs has been implemented to capture the spatial characteristics of this complex river network. In the network hierarchy, predicted discharge from upstream (or lower order stream) networks is used as inputs (in addition to climatic variables) to downstream (higher order) networks. A semi- variogram analysis is used to estimate the temporal lag between input and output variables. The methods are implemented on the 2,704 km2 Winooski River basin in Vermont. Discharge records from six USGS stream gage stations and eight weather stations within the basin provide the training, cross-validation and prediction datasets for this application. Initial predictive results indicate the ANN methods compare well with those published in the literature and outperform traditional methods.