Prediction and Inference of Flow Duration Curves Using Multioutput Neural Networks. Issue 8 (19th August 2019)

• Record Type: Journal Article
• Authors: Worland, Scott. C.; Steinschneider, Scott; Asquith, William; Knight, Rodney; Wieczorek, Michael
• Is Part Of: Water resources research. Volume 55:Issue 8(2019)
• Page Start: 6850
• Page End: 6868
• Publication Date: 2019-08-19
• Subjects: neural network -- flow‐duration curve -- machine learning -- PUB -- hydrology -- streamflow; Hydrology -- Periodicals; 333.91
• DOI: 10.1029/2018WR024463
• Languages: English
• ISSNs: 0043-1397

Abstract

Abstract: We develop multioutput neural network models to predict flow‐duration curves (FDCs) in 9, 203 ungaged locations in the Southeastern United States for six decades between 1950 and 2009. The model architecture contains multiple response variables in the output layer that correspond to individual quantiles along the FDC. During training, predictions are made for each quantile, and a combined loss function is used for back propagation and parameter updating. The loss function accounts for the covariance between the quantiles and generates physically consistent outputs (i.e., monotonically increasing quantiles with increasing nonexceedance probabilities). We use neural network dropout to generate posterior‐predictive distributions for FDCs and test model performance under cross validation. Finally, we demonstrate how local surrogate models, via the Local Interpretable Model‐agnostic Explanations method, can be used to infer the relation between basin characteristics and the predicted FDCs. Results suggest that multioutput neural network models can learn the monotonic relations between adjacent quantiles on an FDC; they result in better predictions than single‐output neural network models that predict each quantile independently, and basin characteristics are most useful for predicting smaller quantiles, whereas bias terms from neighboring quantiles are most informative for predicting higher quantiles. Key Points: Multioutput neural networks (MNNs) generate monotonically increasing flow‐duration curves Monte Carlo dropout captures uncertainty for estimating streamflow quantiles Local surrogate models approximate how the MNN is using basin characteristics for each observation