Keywords
flood prediction, fuzzy modelling, time series knowledge mining, process identification
Start Date
1-7-2008 12:00 AM
Abstract
The challenge of modeling rainfall-runoff processes is to define a suitable functional relationship between input variables representing precipitation measurements and the output runoff. Depending on the system’s state, the amount of precipitation resulting in direct flow varies strongly. Thus, state variables indicating the system’s actual state can enhance the accuracy of rainfall-runoff models significantly. Fuzzy models of Takagi-Sugeno-type are one of the effective approaches to simulate rainfall-runoff processes regarding the different system states. The design of a fuzzy rainfall-runoff model consists of two essential steps: the definition of the structure (input quantities, state variables, rules) and the identification of parameters in the conclusion. The latter one can be solved automatically using data-driven techniques in an optimal way concerning root mean square deviation. However, to solve the task of defining the structure, expert knowledge is mandatory to identify those time series that can effectively be used in a fuzzy model. This expert knowledge is not always available and not necessarily complete or correct. To identify effective state variables semi-automatically, the method Time Series Knowledge Mining (TSKM) has been used. TSKM discovers patterns representing the temporal concepts of duration, coincidence and order. Especially the patterns representing coincidence are valuable as the temporal concept of coincidence is used in fuzzy premises, too: the values of several state variables are evaluated simultaneously to determine the system’s state. TSKM was applied to identify state variables with data from a 7 km2 catchment in the northern Black Forest in Germany. From a set of more than 100 time series that were measured resulting in a huge set of possible state variable configurations, two soil moisture time series were identified. The fuzzy models generated using these two state variables were more efficient than all other models previously generated. Additionally, periods of snowfall and snowmelt could be reliably identified.
Identifying State Variables in Multivariate Hydrological Time Series Using Time Series Knowledge Mining
The challenge of modeling rainfall-runoff processes is to define a suitable functional relationship between input variables representing precipitation measurements and the output runoff. Depending on the system’s state, the amount of precipitation resulting in direct flow varies strongly. Thus, state variables indicating the system’s actual state can enhance the accuracy of rainfall-runoff models significantly. Fuzzy models of Takagi-Sugeno-type are one of the effective approaches to simulate rainfall-runoff processes regarding the different system states. The design of a fuzzy rainfall-runoff model consists of two essential steps: the definition of the structure (input quantities, state variables, rules) and the identification of parameters in the conclusion. The latter one can be solved automatically using data-driven techniques in an optimal way concerning root mean square deviation. However, to solve the task of defining the structure, expert knowledge is mandatory to identify those time series that can effectively be used in a fuzzy model. This expert knowledge is not always available and not necessarily complete or correct. To identify effective state variables semi-automatically, the method Time Series Knowledge Mining (TSKM) has been used. TSKM discovers patterns representing the temporal concepts of duration, coincidence and order. Especially the patterns representing coincidence are valuable as the temporal concept of coincidence is used in fuzzy premises, too: the values of several state variables are evaluated simultaneously to determine the system’s state. TSKM was applied to identify state variables with data from a 7 km2 catchment in the northern Black Forest in Germany. From a set of more than 100 time series that were measured resulting in a huge set of possible state variable configurations, two soil moisture time series were identified. The fuzzy models generated using these two state variables were more efficient than all other models previously generated. Additionally, periods of snowfall and snowmelt could be reliably identified.
