Development and application of a decision group Back-Propagation Neural Network for flood forecasting

SummaryApplying a deterministic model to a system simulation often requires an assumption that the system input–output relationship should hold the whole event. However, this assumption does not apply for nonlinear time-variant systems which possible different outputs can be found due to the various input structures. In this study, a Decision Group Back-Propagation Network (DGBPN) is proposed to avoid the prediction risks inherent in deterministic Back-Propagation Neural Network (BPN) models. The DGBPN was combined with many BPN models. The number of BPN models was dependant on the length of the learning data. The operation process of the DGBPN allows each given input vector of model learning to stand for the scheme of the vector space where it is located, and then uses the output estimation ability to select a suitable BPN model for each given input vector scheme from the BPN model candidates as a representative BPN model (also a member of a decision group). The similarity of input vectors between the forecasting pattern (verification pattern) and model learning patterns was employed to select a BPN model from the DGBPN for forecasting. A flood hydrograph forecast was completed using this method. The model developed in this study not only provides the flexibility necessary for a deterministic BPN model to cope with the volatility of the rainfall–runoff process, but also improves adaptability to reduce estimation error when a system input vector over the range of modeling is encountered. The study concludes with implications for theory, research, and practice. The DGBPN performed well and produced fair forecasting results in verification tests at the Wu-Shi watershed in Taiwan.