Advance flood forecasting for flood stricken Bangladesh with a fuzzy reasoning method

An artificial Neural Network (NN) was successfully applied, in an earlier s tudy, as a prediction tool to forecast water level at Dhaka (Bangladesh), f or up to seven lead days in advance, with a high accuracy level. In additio n, this high accuracy degree was accompanied with a very short computationa l time. Both make NN a desirable advance warming forecasting tool. In a lat er study, a sensitivity analysis was also performed to retain only the most sensitive gauging stations for the Dhaka station. The resulting reduction of gauging stations insignificantly affects the prediction accuracy level. The work concerning the possibility of measurement failure in any of the ga uging stations during the critical flow lever at Dhaka requires prediction tools which can interpret linguistic assessment of flow levels. A fuzzy log ic approach is introduced with two or three membership functions, depending on necessity, for the input stations with five membership functions for th e output station. Membership functions for each station are derived from th eir respective water revel frequency distributions, after the Kohonen neura l network is used to group the data into clusters. The proposed approach in deriving membership function shows a number of advances over the approach commonly used. When prediction results are compared with measured data, the prediction accuracy level is comparable with that of the data driven neura l network approach. Copyright (C) 2000 John Wiley & Sons, Ltd.