Comparison of three alternative ANN designs for monthly rainfall-runoff simulation

The performance of three artificial neural network (ANN) designs that account differently for the effects of seasonal rainfall and runoff variations were investigated for monthly rainfall-runoff simulation on an 815 km(2) watershed in central Oklahoma. The ANN design that accounted explicitly for seasonal variations of rainfall and runoff performed best by all performance measures. Explicit representation of seasonal variations was achieved by use of a separate ANN for each calendar month. For the three ANN designs tested, a regression of simulated versus measured runoff displayed a slope slightly under 1 and positive intercept, pointing to a tendency of the ANN to underpredict high and overpredict low runoff values.