Changes in Water Resources Systems: Methodologies to Maintain Water Security and Ensure Integrated Management (Proceedings of Symposium HS3006 at IUGG2007, Perugia, July 2007).  IAHS Publ. 315, 2007, 228-235.

 

Flood forecasting by Coupling Cluster method and Artificial Neural Networks

 

YIN XIONGRUI1, XIA JUN2 & ZHANG XIANG1

1 State Key Laboratory of Water Resources and Hydropower Engineering Sciences, Wuhan University, Wuhan 430072, China 

aw_yin@163.com

2 Key Laboratory of Water Cycle and Related land Surface Processes, IGSNRR, Chinese Academy of Sciences, Beijing 100101,China

 

Abstract Flood forecasting takes a vital role for flood control and water resources management of catchments. However, it is generally accepted that the relationship of rainfall and runoff is highly complicated, and for a basin, the underlying mechanisms of streamflow generation in rain periods are quite different from those in non-rain periods. The flow hydrograph is broken into several segments, then the rainfall–runoff relationship is separately establish­ed. In this study, we employ two methods to divide the flow hydrograph into several segments. One is the Fuzzy C Means (FCM) method, and the other is the Self-Organizing Feature Map (SOFM). Based on the two clustering results, multi-layer Feedforward Networks (MFN) was used to simulate the rainfall–runoff relationship of each segment. In this way two hybrid artificial neural networks (FCMMFN and SOMMFN) are established. The methods mentioned above are applied to Wangjiachang Reservoir inflow forecasting, in the Hunan province of China, for three-hour-ahead flood forecasting. Forty-five historical flood processes from 11 years (1982–1992) are applied for calibration whilst 14 flood processes from 3 years (1994–1996) are utilized for validation. Antecedent precipitation and streamflow data is input into FCM and SOFM for flow hydrograph decomposing and clustering. The result shows that FCM and SOFM are both able to find the potential knowledge of flow, and that it is easy to find that flow hydrographs as corresponding output is classified into four different stages: (1) low flow; (2) rising flow; (3) flood peak; and (4) recession. Then, for each segment, a MFN is applied to simulate its rainfall–runoff relationship. Results show FCMMFN and SOMMFN are both superior to MFN, i.e. the two hybrid models can simulate precisely the rainfall–runoff relationship simultaneity in low flow, middle flow and high flow. Moreover, FCMMFN and SOMMFN are investigated and compared, and FCMMFN appears to be better.

 

Key words  Artificial Neural Networks; flood forecasting; Fuzzy C Mean; Self-Organizing Feature Map