Development of artificial neural network models for hydrologic prediction in an agricultural watershed.

摘要

In this study, various Artificial Neural Network (ANN) models such as MultiLayer Perceptron (MLP), Radial Basis Function Neural Network (RBFNN) and Long Short Term Memory Recurrent Neural Network (LSTM-RNN) were developed to predict rainfall-runoff process at multiple gauging stations in an agricultural watershed. Land use distribution such as percentage of agriculture, forest, urban, transitional, and water for each subwatershed, daily precipitation (inches) and curve number (CN) were used as input variables to predict daily flow (m3/s) at four USGS gauging stations from 2002-2004 in Eucha Watershed.; Results indicate that LSTM-RNN model was found to be the best model for simulating rainfall-runoff response in this watershed. The LSTM-RNN models have ability to handle time series flow data better than the traditionally used MLP and RBFNN models. A comparative study of ANN and SWAT models indicate that both models are useful tool in forecasting the hydrologic response in agricultural watershed. Overall, ANN should be regarded as an alternative to more traditional rainfall-runoff methods and the model selection should depend on the availability of data and modeling objective at multiple gauging stations.; The driving force of non-point source (NPS) pollution is the rainfall-runoff process, which is a complex interaction between the components of the water cycle and geographic and land use factors. Prediction of hydrologic parameter is usually done through a combination of extensive environmental monitoring and computer modeling. This study presents a procedure for developing ANN models to predict hydrologic response of agricultural watersheds at multiple gauging stations.