Development of a distributed artificial neural network for hydrologic modeling

Hydrological models are used to represent the rainfall-runoff and pollutant transport mechanisms within watersheds. Accurate representation of these dynamic and complex natural processes within a watershed is an important step in managing and protecting a watershed. Artificial neural network (ANN) models are often used in hydrologic modeling. Typical ANN models are trained to use lumped data. However, watershed characteristics used as inputs in hydrological modeling are spatially and often temporally dynamic. Therefore, a lumped model does not have the ability to represent changes in spatial dynamics of a watershed. Therefore, the purpose of this study was to develop and test a distributed ANN model for simulating the rainfall-runoff process in the L'Anguille River Watershed located in Eastern Arkansas. The watershed was divided into nine sub-basins to account for the spatial dynamics of flow within the watershed. Inputs for the model were rainfall, average temperature, antecedent flow and curve number. Output was runoff, collected from gage-stations at Colt and Palestine representing two of the sub-basins. Daily SCS curve numbers were developed and adjusted for crop planting and harvesting dates and crop rotation practices in each sub-basin. The model had nine layers with one neuron each to represent the nine sub-basins. The layers were connected so that if one sub-basin spatially flowed into another, its output would be an input for the downstream sub-basin. The model performed well, showing R ² values of 0.93 and 0.98 and Nash-Sutcliffe Efficiency values of 0.92 and 0.97 for the validation and test datasets.