Keywords
Hybrid models; flood inundation models; emulation models.
Start Date
17-9-2020 2:40 PM
End Date
17-9-2020 3:00 PM
Abstract
Flood management often relies on flood inundation information provided by 2D hydrodynamic models. These models are computationally expensive, which prevents their use in applications involving a large number of model runs, such as uncertainty analysis or ensemble forecasting. Recently, machine-learning techniques, such as artificial neural networks (ANNs), have been used to develop fast emulation models for flood inundation modelling, where the authors have found that ANNs can significantly improve modelling efficiency and produce reasonable flood inundation predictions when there are sufficient data. However, the performance of these ANN models is limited in data-sparse regions, which is expected for any data-driven approach. In this study, we propose a hybrid modelling approach based on ANNs to improve model performance in data-sparse regions. In the hybrid modelling approach, the valuable information in data-rich regions is used to assist the development of models in data-sparse regions based on the inherent relationships between these regions. The proposed hybrid modelling approach is applied with three different ANNs with different model complexities to a real-world case study in Australia, and compared with the ANN-only modelling approach in terms of their predictive performance. The comparison results show that the hybrid modelling approach can improve model performance in data-sparse regions significantly, and the hybrid model based on the most complex ANN performed the best. However, the improvement in model performance is lost in the regions with relatively more data, where ANN models directly mapping the input-output relationships performed better. This study highlights the importance of considering the trade-offs between model complexity and data availability in the development of machine-learning models and demonstrates the capability of a hybrid modelling approach in improving model performance in data-sparse regions by leveraging valuable information in data-rich regions.
A Machine-Learning based Hybrid Modelling Approach for Flood Inundation Modelling
Flood management often relies on flood inundation information provided by 2D hydrodynamic models. These models are computationally expensive, which prevents their use in applications involving a large number of model runs, such as uncertainty analysis or ensemble forecasting. Recently, machine-learning techniques, such as artificial neural networks (ANNs), have been used to develop fast emulation models for flood inundation modelling, where the authors have found that ANNs can significantly improve modelling efficiency and produce reasonable flood inundation predictions when there are sufficient data. However, the performance of these ANN models is limited in data-sparse regions, which is expected for any data-driven approach. In this study, we propose a hybrid modelling approach based on ANNs to improve model performance in data-sparse regions. In the hybrid modelling approach, the valuable information in data-rich regions is used to assist the development of models in data-sparse regions based on the inherent relationships between these regions. The proposed hybrid modelling approach is applied with three different ANNs with different model complexities to a real-world case study in Australia, and compared with the ANN-only modelling approach in terms of their predictive performance. The comparison results show that the hybrid modelling approach can improve model performance in data-sparse regions significantly, and the hybrid model based on the most complex ANN performed the best. However, the improvement in model performance is lost in the regions with relatively more data, where ANN models directly mapping the input-output relationships performed better. This study highlights the importance of considering the trade-offs between model complexity and data availability in the development of machine-learning models and demonstrates the capability of a hybrid modelling approach in improving model performance in data-sparse regions by leveraging valuable information in data-rich regions.
Stream and Session
false