1st Edition
Hybrid models for Hydrological Forecasting: integration of data-driven and conceptual modelling techniques UNESCO-IHE PhD Thesis
This book presents the investigation of possibilities and different architectures of integrating hydrological knowledge and conceptual models with data-driven models for the purpose of hydrological flow forecasting. Models resulting from such integration are referred to as hybrid models. The book addresses the following specific topics:
A classification of different hybrid modelling approaches in the context of flow forecasting.
The methodological development and application of modular models based on clustering and baseflow empirical formulations.
The integration of hydrological conceptual models with neural network error corrector models and the use of committee models for daily streamflow forecasting.
The application of modular modelling and fuzzy committee models to the problem of downscaling weather information for hydrological forecasting.
The results of this research show the increased forecasting accuracy when modular models, which integrate conceptual and data-driven models, are considered. Committee machine modelling show to be able to manage increased lead time with an acceptable accuracy.
Summary
1 Introduction
1.1 Background
1.2 Flood management and forecasting
1.2.1 Flood management measures
1.2.2 Operational flow forecasting
1.3 Hydrological models
1.3.1 Classification
1.3.2 HBV process-based model
1.4 Data-driven models
1.5 Objectives of the research
1.6 Terminology
1.7 Outline
2 Framework for hybrid modeling
2.1 Introduction
2.2 General considerations and assumptions
2.3 Hybrid modelling framework
2.3.1 Classification of hybrid models
2.3.2 Relationships between model classes
2.4 Committee machines and modular models
2.5 Measuring model performance
2.6 Discussion and conclusions
3 Optimal modularization of data-driven models
3.1 Introduction
3.2 Methodology of modular modelling
3.3 Modularization using clustering (MM1)
3.4 Modularization using sub-process identification (MM2)
3.5 Modularization using time-based partitioning (MM3)
3.6 Modularization using spatial-based partitioning
3.7 Optimal combination of modularization schemes
3.8 Conclusions
4 Building data-driven hydrological models: data issues
4.1 Introduction
4.2 Case study (Ourthe river basin - Belgium)
4.3 Procedure of data-driven modelling
4.4 Preparing data and building a model
4.5 The problem of input variables selection
4.5.1 Inputs selection based on correlation analysis
4.5.2 Selection based on Average Mutual Information (AMI)
4.6 Influence of data partitioning
4.7 Influence of ANN weight initialization
4.7.1 Models not using past discharges as inputs (RR)
4.7.2 Models using past discharges as inputs (RRQ)
4.8 Various measures of model error
4.9 Comparing the various types of models
4.10 Discussion and conclusions
5 Time and process based modularization
5.1 Introduction
5.2 Catchment descriptions
5.3 Input variable selection
5.4 Comparison to benchmark models
5.5 Modelling process
5.6 Results and discussion
5.7 Conclusions
6 Spatial-based hybrid modelling
6.1 Introduction
6.2 HBV-M model for Meuse river basin
6.2.1 Characterisation of the Meuse River basin
6.2.2 Data validation
6.3 Methodology
6.3.1 HBV-M model setup
6.3.2 Scheme 1: Sub-basin model replacement
6.3.3 Scheme 2: Integration of sub-basin models
6.4 Application of Scheme 1
6.4.1 Inputs selection and data preparation for DDMs
6.4.2 Data-driven sub-basin models
6.4.3 Analysis of HBV-S simulation errors
6.4.4 Replacements of sub-basin models by ANNs
6.5 Application of Scheme 2
6.6 Discussion
6.6.1 Scheme 1
6.6.2 Scheme 2
6.7 Conclusions
7 Hybrid parallel and sequential models
7.1 Introduction
7.2 Metodology and models setup
7.2.1 Meuse river basin data and HBV model
7.2.2 ANN model setup
7.3 Data assimilation (error correction)
7.4 Committee and ensemble models
7.5 Forecasting scenario
7.6 Results and discussion
7.6.1 Single forecast results
7.6.2 Results on multi step forecast
7.7 Conclusions
8 Downscaling with modular models
8.1 Introduction
8.2 Fuzzy committee
8.3 Case study: Beles River Basin, Ethiopia
8.4 Beles River Basin
8.5 Methodology
8.5.1 ANN model setup
8.5.2 Committee and modular models
8.5.3 Fuzzy committee machine
8.6 Results
8.7 Conclusions
9 Conclusions and Recommendations
9.1 Hybrid modelling
9.2 Modular modelling
9.3 Downscaling with modular models
9.4 Parallel and serial modelling architectures
9.5 Data-driven modelling
9.6 Conclusion in brief
Bibliography
A State-Space to input-output transformation
A.1 State-space and input-output models
B Data-driven Models
B.1 Arti¯cial Neural Networks (Multi-layer perceptron)
B.2 Model Trees (M5P)
B.3 Support Vector Machines
C Hourly forecast models in the Meuse
C.1 Methodology
C.2 Neural network model (ANN)
C.3 Results
List of Figures
List of Tables
List of acronyms
Samenvatting
Acknowledgements
About the author
Biography
Gerald Corzo received his degree of civil engineering from the Escuela Colombian de Ingenieria, Faculty of Engineering (Bogotá, Colombia). He joined the Universidad Francisco de Paula Santander (UFPS), as a lecturer in Mathematics and Statistics (2000-2003). After graduating with the highest degree in his course he moved to Bogotá and served as lecturer in different universities. In October 2003 he moved to Delft in the Netherlands, there he joined the Hydroinformatics Master program at the UNESCO-IHE Institute for Water Education. He was awarded Master of Science in May 2005. His work explored the use of rule-based modelling and Committee of data-driven models in hydrological forecasting. In his thesis he showed a modular scheme approaching the incorporation of hydrological knowledge in a committee model. He continued with his work in the Phd, where he extended the analysis of artificial intelligence methods applied to hydrological phenomena.
