Forecasting Daily Streamflow Discharges Using Various Neural Network Models and Training Algorithms

Streamflow forecasting based on past records is an important issue in both hydrologic engineering and hydropower reservoir management. In the study, three artificial Neural Network (NN) models, namely NN with well-known multi-layer perceptron (MLPNN), NN with principal component analyses (PCA-NN), and NN with time lagged recurrent (TLR-NN), were used to 1, 3, 5, 7, and 14 ahead of daily streamflow forecast. Daily flow discharges of Haldizen River, located in the Eastern Black Sea Region, Turkey the time period of 1998-2009 was used to forecast discharges. Backpropagation (BP), Conjugate Gradient (CG), and Levenberg-Marquardt (LM) were applied to the models as training algorithm. The result demonstrated that, firstly, the forecast ability of CG algorithm much better than BP and LM algorithms in the models; secondly, the best performance was obtained by PCA-NN and MLP-NN for short time (1, 3, and 5 day-ahead) forecast and TLR-NN for long time (7 and 14 day-ahead) forecast.