Reliability Analysis of Flood Stage Forecasting using Neural Networks Medel I. Model Development and Application

Elman Discrete Recurrent Neural Networks Model (EDRNNM), one of special type neural networks model, was developed to be a highly suitable tool producing a very high degree of flood stage forecasting accuracy at Musung station (No. 1) of Wi-stream catchment, one of IHP representative basins in South Korea. A relative new approach, EDRNNM, has recurrent feed-back nodes and virtual small memory on the networks architecture. The 135 different training patterns, which Involve hidden nodes, standardization method, data length and lead-time, were selected to minimize the architectural uncertainty during training performance. The model parameters, optimal connection weights and biases, were estimated during training performance and they were applied to evaluate model validation performance. A cross-validation method was applied to select the best patterns of 6 different training patterns during validation performance. FeedForward Neural Networks Model (FFNNM), one of traditional ANNs-based models, was introduced to compare with the results of EDRNNM performance. From model training and validation performance, EDRNNM was proved to be outstanding model for the flood stage forecasting in the Wi-stream catchment. Even if, however, much of the unknown uncertainties were eliminated during training and validation performance, some of them still remain on input data information. Therefore, the continuous research was required to reduce useless input data during investigation of the importance of each of the Input data for reliability analysis.