PREDICTION OF 2-YEAR PEAK STREAM DISCHARGES USING NEURAL NETWORKS

The cascade correlation neural network was used to predict the two-yea r peak discharge (Q(2)) for major regional river basins of the contine ntal United States (US). Watersheds ranged in size by four orders of m agnitude. Results of the neural network predictions ranged from correl ations of 0.73 for 104 test data in the Souris-Red Rainy river basin t o 0.95 for 141 test data in California. These results are improvements over previous multilinear regressions involving more variables that s howed correlations ranging from 0.26 to 0.94. Results are presented fo r neural networks trained and tested on drainage area, average annual precipitation, and mean basin elevation. A neural network trained on r egional scale data in the Texas Gulf was comparable to previous estima tes of Q(2) by regression. Our research shows Q(2) was difficult to pr edict for the Souris-Red Rainy, Missouri, and Rio Grande river basins compared to the rest of the US, and acceptable predictions could be ma de using only mean basin elevation and drainage areas of watersheds.