Adjusting for Conditional Bias in Process Model Simulations of Hydrological Extremes: An Experiment Using the North Wyke Farm Platform.

Stelian Curceac,Lianhai Wu,Alice Milne,Peter M Atkinson,Paul Harris

doi:10.3389/frai.2020.565859

Abstract

Peak flow events can lead to flooding which can have negative impacts on human life and ecosystem services. Therefore, accurate forecasting of such peak flows is important. Physically-based process models are commonly used to simulate water flow, but they often under-predict peak events (i.e., are conditionally biased), undermining their suitability for use in flood forecasting. In this research, we explored methods to increase the accuracy of peak flow simulations from a process-based model by combining the model’s output with: a) a semi-parametric conditional extreme model and b) an extreme learning machine model. The proposed 3-model hybrid approach was evaluated using fine temporal resolution water flow data from a sub-catchment of the North Wyke Farm Platform, a grassland research station in south-west England, United Kingdom. The hybrid model was assessed objectively against its simpler constituent models using a jackknife evaluation procedure with several error and agreement indices. The proposed hybrid approach was better able to capture the dynamics of the flow process and, thereby, increase prediction accuracy of the peak flow events.

Highlights

In the United Kingdom, the estimated yearly cost of damages caused by floods is over £1 billion (Collet et al, 2017)
We explored combining statistical and machine learning techniques with flow simulations obtained from a process-based models (PBMs) to increase the accuracy of forecasting peak flow events
We present a general description of the conditional extreme model (CEM) (Heffernan and Tawn, 2004) and the extreme learning machine (ELM) (Huang et al, 2006) and explains how they can be applied to peak flow events obtained from a chosen PBM in a hybrid context

Summary

Introduction

In the United Kingdom, the estimated yearly cost of damages caused by floods is over £1 billion (Collet et al, 2017). Accurate and reliable forecasting of extreme flow events is crucial for planning and implementing measures to mitigate their effects and so protect lives, properties and services. The magnitude and frequency of floods is likely to increase as a result of climate change (Kundzewicz et al, 2007; Bates et al, 2008; Field et al, 2012) and this could push ecosystems beyond the threshold of normal disturbance (Thibault and Brown, 2008). Increased runoff and flooding intensify erosion and result in higher sediment and nutrient losses that can lead to soil degradation and high concentrations of pollutants in water courses (Bouraoui et al, 2004). Different approaches have been proposed for more accurate modeling and forecasting of peak flows with reduced uncertainty.

Methods

Results

Discussion

Conclusion