004 Datenverarbeitung; Informatik
Refine
Document Type
- Article (1)
- Doctoral Thesis (1)
Language
- English (2)
Is part of the Bibliography
- yes (2)
Keywords
- modelling (2) (remove)
We systematically explore the effect of calibration data length on the performance of a conceptual hydrological model, GR4H, in comparison to two Artificial Neural Network (ANN) architectures: Long Short-Term Memory Networks (LSTM) and Gated Recurrent Units (GRU), which have just recently been introduced to the field of hydrology. We implemented a case study for six river basins across the contiguous United States, with 25 years of meteorological and discharge data. Nine years were reserved for independent validation; two years were used as a warm-up period, one year for each of the calibration and validation periods, respectively; from the remaining 14 years, we sampled increasing amounts of data for model calibration, and found pronounced differences in model performance. While GR4H required less data to converge, LSTM and GRU caught up at a remarkable rate, considering their number of parameters. Also, LSTM and GRU exhibited the higher calibration instability in comparison to GR4H. These findings confirm the potential of modern deep-learning architectures in rainfall runoff modelling, but also highlight the noticeable differences between them in regard to the effect of calibration data length.
Systems of Systems (SoS) have received a lot of attention recently. In this thesis we will focus on SoS that are built atop the techniques of Service-Oriented Architectures and thus combine the benefits and challenges of both paradigms. For this thesis we will understand SoS as ensembles of single autonomous systems that are integrated to a larger system, the SoS. The interesting fact about these systems is that the previously isolated systems are still maintained, improved and developed on their own. Structural dynamics is an issue in SoS, as at every point in time systems can join and leave the ensemble. This and the fact that the cooperation among the constituent systems is not necessarily observable means that we will consider these systems as open systems. Of course, the system has a clear boundary at each point in time, but this can only be identified by halting the complete SoS. However, halting a system of that size is practically impossible. Often SoS are combinations of software systems and physical systems. Hence a failure in the software system can have a serious physical impact what makes an SoS of this kind easily a safety-critical system. The contribution of this thesis is a modelling approach that extends OMG's SoaML and basically relies on collaborations and roles as an abstraction layer above the components. This will allow us to describe SoS at an architectural level. We will also give a formal semantics for our modelling approach which employs hybrid graph-transformation systems. The modelling approach is accompanied by a modular verification scheme that will be able to cope with the complexity constraints implied by the SoS' structural dynamics and size. Building such autonomous systems as SoS without evolution at the architectural level --- i. e. adding and removing of components and services --- is inadequate. Therefore our approach directly supports the modelling and verification of evolution.