Multistep Ahead Water Level Forecasting using Different Artificial Neural Network Training Algorithms

Jianjun Yu, Xiaosheng Qin, and Ole Larsen


Water level forecasting, artificial neural networks, fuzzy c-means, clustering


Many artificial neural network (ANN) models including global and data-partition (e.g. clustering) based models have been developed and successfully applied to studies of hydrologic forecasting. There is a need to compare the applicability of the global ANN models trained with different algorithms and clustering-based ANN models in hydrologic forecasting. This study presented a comparison of 8 ANN models trained with standard back-propagation (BP), gradient descent with adaptive learning rate (GDA), resilient back-propagation (RP), conjugate gradient with Fletcher-Reeves update (CGF), conjugate gradient with Polak-RibiƩre update (CGP), Broyden-Fletcher-Goldfarb-Shanno (BFGS), one step secant (OSS), and Levenberg-Marquardt (LM) algorithms as well as one fuzzy c-means (FCM) clustering based ANN model for 1- to 5-day ahead water-level forecasting in Heshui catchment, China. The results showed that BFGS and LM were the best training algorithms in all prediction scenarios. The CGP, CGF, OSS and RP algorithms performed better than BP and GDA. FCM-based model showed a satisfactory result but performed no better than BFGS or LM trained models.

Important Links:

Go Back