Development of hybrid wavelet-ANN model for hourly flood stage forecasting

Abstract

Hourly flood forecasting with sufficient lead time is necessary to develop preparedness and facilitate quick response in adopting appropriate flood mitigation plans and rehabilitation actions. This study explores the potential of hybrid wavelet-ANN (wavelet-Artificial neural network) to model flood events in the Achankovil river basin in Kerala. Hourly water level at the Konni gauging site is predicted with 1-, 3- and 6-h lead times. The hybrid wavelet-ANN approach has proved its capabilities in hydrological studies due to the capability of wavelet in multi-resolution analysis and its ability to deal with non-stationarities. The three hybrid wavelet-ANN models developed were trained considering various combinations of flood events. The significant antecedent water level and precipitation values were selected based on statistical analysis and a study of the hydrology of the catchment. Decomposition of the input signals using wavelet was performed by selecting a suitable mother wavelet and an appropriate decomposition level. Results of this study clearly demonstrate the capability of the hybrid models to forecast floods satisfactorily up to 6 h lead time. The capability of the hybrid model is proved by comparison with a simple ANN model. The WANN models developed were able to predict the magnitude of peak value and the time to peak value more accurately when compared to the ANN models. These two parameters of a flood hydrograph are very important in many water resources applications.