The Huai River is one of the seven largest rivers in China. Due to the high density of the population, the dense communication networks and the economic role, it is a very important region of the country.
The Huai River lies in the climatical transitional zone from the northern to southern China. To the north of the Huai River is the South Temperate Zone and to the south of the Huai River is the North Subtropical Zone. Under the dominating monsoon climate, precipitation and evaporation change significantly both in time and space. Approximately once in every 5 years the Huai River Basin is seriously threatened and damaged by heavy floods. The dikes along the Huai River have been designed on basis of a flood water level with a return period of 20 years. Hence, it is very helpful to build rainfall-runoff model for flood loss mitigation.
Usually there are two approaches to rainfall-runoff modelling: 1) physically-based hydrological modelling using mathematical models that solve set of equations describing the physical hydrological process and 2) data-driven models based on establishing input-output non-linear relationships between the rainfall and runoff analyzing long time-series of observable. In this study, second approach for building a rainfall-runoff model is investigated.
Artificial Neural Networks (ANNs) has been successfully applied to rainfall-runoff modelling in the past. In this study, three types of ANNs, i.e. MLPs, recurrent and modular, are investigated. Generally the MLPs can simulate the relationships between rainfall and runoff. For single flood event, it can simulate the event well, both for the low discharges and peak discharges (>500m3/s), but for complicated events (more than 1 flood peaks occur consequently), the simulated results of low discharges are generally good whereas the results of peak discharges are not satisfactory. Recurrent and modular type of networks were successfully applied in this study to setup rainfall-runoff models to simulate the hydrological behavior of the study area in the Huai River as well as MLPs.
The linear activation function was found necessary for the output nodes of the networks. The use of Tanh function was to enable nonlinearity of the network. However it was not adopted for the output nodes because the tangential function forces an output variable to be scaled by a known maximum value.
Spatial distribution of the rainfall was found to have important impact for the peak discharges simulation, however a proper selection of the spatial rainfall information must be investigated. Architecture with too many input nodes and limited amount of training data may introduce 'noise' to the network and the network does not work well.
Selection of a proper lag time for the network was found to influence the simulation results; the proper lag time should be selected depending on the investigation and available data.
A brute-force search for an optimal combination between the spatial distribution of the rainfall, proper lag time and proper architecture of the ANNs, was successfully applied in this study in order to build a rainfall-runoff model.
Generally the objective of this study has been achieved, but more investigation on the spatial-temporal distribution of the rainfall as an input patterns and application of mixture of models (hybrid modelling) should be tested.
Back to the list of MSc abstracts