Author(s): Zongxue Xu; Kuniyoshi Takeuchi; Hiroshi Ishidaira

Linked Author(s):

Keywords: Flood; Reservoir; Real-time forecasting; Artificial neural network

Abstract: An artificial neural network (ANN), which emulates the parallel distributed processing of the human nervous system, is a flexible mathematical model that is capable of simulating non-linear physical processes. It has been proven particularly efficient for problems in which the internal characteristics of the processes are difficult to be described by physical equations such as hydrologic processes. The study presented in this paper has investigated the capability of the ANN technique for real-time flood forecasting. Two case studies are presented, in which ANN models are employed to forecast 1-to 6-hour ahead inflows into Ikusaka reservoir, and 1-to 3-hour ahead discharge and water level along the Tsurumi River in Japan. The ANN models are trained to remember historical storm patterns for reproduction of relevant pattern for new floods. It was found that the ANN model might provide quite accurate forecasting when an optimum number of spatial and temporal inputs are included in the network, and the network with shorter lead-time produced better performance. Because the ANN model has no physically realistic parameters, it is not a substitute for conceptual watershed models. However, the ANN technique does provide an effective alternative for real-time flood forecasting when the internal characteristics of the watershed are complex or difficult to be simulated.

DOI:

Year: 2002