Uncertainty Analysis in Rainfall-Runoff Modelling: Application of Machine Learning Techniques

Rainfall-runoff models are widely used in hydrology for a large range of applications and play an important role in planning and management of water resources. In practice engineering decisions are often based on a single model run as if the model is perfect and input data is error free. The predictions made by such a model are inherently uncertain and therefore, it is vital that uncertainty should be recognized and properly accounted for. Because of recent global warming, uncertainty analysis in hydrological modelling is becoming more important.

A number of uncertainty analysis methods have been developed and successfully used in the past. The novelty of this PhD research is in using powerful machine learning techniques to build predictive models of uncertainty with application to hydrological models.Two different methods were developed and tested. The first one focuses on parameter uncertainty analysis by emulating the results of Monte Carlo simulations of hydrological models using efficient machine learning techniques. The second method aims at modelling uncertainty by building an ensemble of specialised machine learning models on the basis of past hydrological model’s performance. Methods employed include artificial neural networks, model trees, locally weighted regression and fuzzy logic. As supporting techniques, evolutionary and other randomised search optimisation methods in multi-criteria context are used.

The developed methods were applied to rainfall-runoff models of three contrasting catchments. This research demonstrates the capacity of machine learning techniques for building accurate and efficient predictive models of uncertainty.