Flood Forecasting and Uncertainty Assessment Using Wavelet- and Bootstrap-Based Neural Networks

1. Introduction

Flood forecasting is the estimation of future water levels or flows at a single or multiple sites of a river system for different lead times. Daily flood forecasts are essential for water resources planning and management including potential water supply for domestic needs, irrigation scheduling, hydropower generation, regulating flows through reservoirs and barrages and for issuing flood warning. Flood warning is essential to mitigate natural disaster by undertaking appropriate evacuation and rehabilitation plans. The necessity of accurate and reliable river flood forecasts is increasingly being felt with increasing demands on water resources and catastrophic disaster due to flood that is continuously increasing due to economic development and demographic expansion.

Classical time series models such as auto regressive integrated moving average (ARIMA) are widely used for hydrological time series forecasting as they are accepted as a standard representation of a stochastic time series (Maier & Dandy, 1997). However, these models are basically linear models which make use of classical statistics to analyse the historical data. To overcome the limitations of classical time series models, a wide variety of rainfall-runoff models have been developed and applied for flood forecasting ranging from complex physically based to simple black box models. Black box models in the form of neural networks (NNs) have been widely used last few decades for flood forecasting and have been accepted as a good alternative to physically based and conceptual models (ASCE, 2000a,b). The ability of NNs in modelling the complex nonlinear relationship between inputs and outputs without explicitly accounting for the physical processes has increased the number of applications in flood forecasting. Most importantly NN models need limited inputs such as water level, discharge, rainfall or sometimes only a single input, whereas physically based models require several additional parameters which are difficult to measure because of temporal and spatial variability. This approach has also been criticised for making models overly complex which lead to problems of over parameterisation and equifinality (Beven, 2006) causing large prediction uncertainty.