10:15 AM - 10:30 AM
[SCG52-06] Prediction of natural disasters based on a hybrid approach of physics-driven and data-driven method
★Invited Papers
Keywords:Prediction of phenomena, data-driven science, spatio-temporal data
Data-driven science methods, which have made remarkable progress in recent years, are expected to be utilized in the field of disaster management, and their applications in hazard prediction, current situation assessment, and disaster response are being studied. On the other hand, because it is difficult to assure the quality and quantity of data under the inherent uncertainty of natural phenomena and the rarity of major events, the application of data-driven science methods is becoming increasingly difficult. In particular, predicting the temporal evolution of disaster events is necessary to predict future states, including situations that have never been experienced before, and improving the accuracy of such extrapolative predictions is a major challenge for data-driven science.
To solve this problem, an approach that combines physical methods where natural laws are modeled with data-driven science methods has recently gained popularity. Such an approach makes complementary use of the advantages of data-driven methods, which build models flexibly from data, and the advantages of physical methods, which enable extrapolative prediction by applying natural laws.
In this presentation, after an overview of the integration of physical and data-driven methods, a method for short-term rainfall prediction is presented as a concrete example of research by the presenter. In this method, the governing equation of rainfall is divided into two terms, i.e., the substantial change term and the advection term under the Lagrangian coordinates, then each term is modeled by different data-driven methods. The substantial change term is modeled using the Koopman mode analysis, a data-driven method for discovering the laws behind dynamic phenomena. The method's performance in representing the development and decay of rainfall is presented compared with conventional short-term forecasting methods.
To solve this problem, an approach that combines physical methods where natural laws are modeled with data-driven science methods has recently gained popularity. Such an approach makes complementary use of the advantages of data-driven methods, which build models flexibly from data, and the advantages of physical methods, which enable extrapolative prediction by applying natural laws.
In this presentation, after an overview of the integration of physical and data-driven methods, a method for short-term rainfall prediction is presented as a concrete example of research by the presenter. In this method, the governing equation of rainfall is divided into two terms, i.e., the substantial change term and the advection term under the Lagrangian coordinates, then each term is modeled by different data-driven methods. The substantial change term is modeled using the Koopman mode analysis, a data-driven method for discovering the laws behind dynamic phenomena. The method's performance in representing the development and decay of rainfall is presented compared with conventional short-term forecasting methods.