Entropy-based neural networks model for flow duration curves at ungauged sites

•AEDI is introduced in ANN to more accurately predict FDC at ungauged sites.•The ANN model was trained and tested on 147 streams in Ontario.•Both location and scale parameters were highly sensitive to MAP, MAS and AEDI.•The location parameter was most sensitive to DA, SHPF, VEG.•The scale parameter was sensitive to DA, SLP, BFI.

SummaryAn apportionment entropy disorder index (AEDI), capturing both temporal and spatial variability of precipitation, was introduced as a new input parameter to an artificial neural networks (ANN) model to more accurately predict flow duration curves (FDCs) at ungauged sites. The ANN model was trained on the randomly selected 2/3 of the dataset of 147 gauged streams in Ontario, and validated on the remaining 1/3. Both location and scale parameters that define the lognormal distribution for the FDCs were highly sensitive to the driving climatic factors, such as, mean annual precipitation, mean annual snowfall, and AEDI. Of the long list of watershed characteristics, the location parameter was most sensitive to drainage area, shape factor and percent area covered by natural vegetation that enhanced evapotranspiration. However, scale parameter was sensitive to drainage area, watershed slope and the baseflow index. Incorporation of the AEDI in the ANN model improved prediction performance of the location and scale parameters by 7% and 21%, respectively. A case study application of the new model for the design of micro-hydropower generators in ungauged rural headwater streams was presented.