A distributed artificial neural network model for watershed-scale rainfall-runoff modeling.

摘要

Distributed models capture the spatiotemporal dynamics of hydrologic processes such as rainfall-runoff, which enable them to be more insightful tools for watershed management. Although artificial neural network (ANN) offers a framework for developing cause-effect relationships, the majority of the ANN-based hydrologic models are lumped. This study explores the ability of an ANN to represent the spatiotemporal dynamics of a watershed process. A distributed ANN (dANN) model was developed using a cascade-forward multi-layer perceptron model framework to represent the RR process. The model was tested for a sample watershed, the L'Anguille River watershed in eastern Arkansas, USA. This model incorporated the known spatial dynamics of rainfall and flow in the watershed through laterally connected blocks that represented subbasins. The model had high predictability, simulated the flow with high accuracy (RMSE=0.15 to 0.37 mm, R2=0.93 to 0.99) at daily time scale when compared to measured flow, and represented the spatial dynamics of flow as well as the SWAT model. The dANN model can be a computationally efficient tool for simulating spatiotemporally dynamic watershed processes, which can expand its applications into water quality modeling, flood forecasting, and watershed management.