Author(s): Dimitri P. Solomatine
Linked Author(s):
Keywords: Streamflow forecasting; Data-driven modeling; Parsimonious models
Abstract: Streamflow forecasting is normally based on physically-based (process) models based on equations describing the water flow. Data-driven models use methods of machine learning (neural networks, M5 model trees, support vector machines, etc. ) and built using the historical data describing flows and hydrological loads. Modular models include components each of which is responsible for a particular hydrological condition, and they are typically more accurate than the overall global models. It appears that the component local models can be made simpler than the overall model; such parsimonious models are better accepted by domain experts. An issue here is optimization of modular models and the inclusion of human experts into the process of model building. The paper addresses these and other issues of building modular models, presents the new algorithms, and reports the results of using them in three hydrological case studies.
Year: 2005