Downstream river flow forecasting by hybrid neural network model with upstream flow and precipitation inputs

Xiao Yun Chen, Kwok Wing Chau

Tuesday 30 june 2015

12:42 - 12:45h at South America (level 0)

Themes: (T) Flood risk management and adaptation, (ST) Flood risk assessment, Poster pitches

Parallel session: Poster pitches: 5J. FloodRisk - Assessment

The need for accurate forecasting of river flow has grown rapidly in the past decades because of its importance in flood risk assessment and coastal engineering. Previous models predicted river flow mainly based on its own time-series data without consideration of the spatial factors. In this paper, river flow and precipitation data from the upstream sections are imposed as critical inputs to a hybrid neural network (HNN) model for downstream river flow forecasting. The HNN model combines fuzzy pattern-recognition concept and continuity equation into neural network, in which the fuzzy-and-nonlinear behavior of river flow and river station represented by storage reservoir are taken into account. In particular, areal precipitation over the entire river basin is computed by Thiessen polygon method and regarded as a potential input variable. The performance of model with areal precipitation is compared to that without precipitation and that with point precipitation at observed station. For the purpose of this study, the Yellow River with four recorded stations in Altamaha River basin of Georgia is selected as a case study site. A detailed comparison of the overall performances indicates that the combined input model (with all upstream flow and areal precipitation inputs) is the best. It is thus concluded that the flows and precipitations in upstream river sections play an indispensable role in the forecasting of downstream river flow in a river basin. The ability of HNN model to elucidate simultaneously both spatial and temporal information with appropriate input is demonstrated and verified as well.