Application of Data-Driven Techniques in Semi-Distributed Hydrological Modelling

Mr. Hidayat

Abstract

River flow forecasting is very important for flood management, navigation, water allocation, etc. Currently, there is a growing need of knowing the quantity of flow several hours or days ahead to anticipate extreme conditions, i.e. flood and drought. Rainfall-runoff modelling allows such forecast to be made.

This thesis deals with improvement of flow forecasting in a meso-scale catchments by incorporating data-driven modelling techniques. It explores how the sub-basin models influence the overall performance of the system. The HBV semi-distributed model of the Meuse (developed by RIZA) is used in this study. Some experiments had been carried out to explore the applicability of data-driven techniques in improving the model performance.

Three main experiments were performed to approach integration options of the data driven models into the Meuse model:

· Analysis of error contribution in the HBV model: The information from measurement stations was incorporated in sequence to see the sensitivity of the overall performance to the distribution and routing scheme used in the model integration and to identify which sub-basin to be improved further. This analysis helped in defining bounds of reasonable maximum improvement that can be obtained with sub-basin model replacement.

· Basin models replacement: This was done by the generation of multiple data models for each sub-basin. Data-driven modelling techniques used were linear regression, model tree and artificial neural network.

· Incorporation of data-driven models into the meso-scale conceptual model: A representation of different simulations using neural networks as inflow into the HBV sub-basin distribution was performed. This was done to incorporate the forecasting ability of the networks with the simulation of the HBV model. To approach an operational forecast process the precipitation information was adapted by the use of different ranges of noise.

Error analysis showed that except for linear regression model, data-driven models perform better than the HBV. These results turned out to be the basis in the integration of data-driven models in HBV model. HBV sub-basin model replacements by ANN models resulted in improvement of the simulated discharge. Assessment of model sensitivity to the presence of noise in precipitation data showed that the HBV+ANN model was less sensitive to disturbance in precipitation data than the original HBV. Experiment on integrating HBV sub-basin local model outputs using ANN resulted in

better overall basin output discharge simulation.

Keywords: data-driven model, semi-distributed hydrological model, discharge, HBV, ANN, river Meuse catchment