1:45 PM - 3:15 PM
[AAS04-P05] Efficient generation of atmospheric forcings for a storm surge model using deep learning
Keywords:Storm surge, Typhoon, Deep learning
A typhoon-induced storm surge model requires 10-m wind and sea level pressure fields as forcings commonly obtained using parametric models or a fully dynamical simulation by numerical weather prediction (NWP) models. Parametric models have been reported to work well for storm surge simulations (e.g., Vijayan et al., 2021). However, poor predictive skills are typically exhibited in areas far from the typhoon center, during the extratropical transition, and landfall due to topographic effects. NWP models are generally more accurate than the parametric models, albeit they are also subjected to several limitations in resolving typhoon intensity. Nevertheless, recent atmospheric and computer science developments have led to substantial improvements in the NWP models. Besides a finer NWP model grid resolution (Dullaart et al., 2020), a data assimilation scheme can be combined with an NWP model to improve typhoon structure resolvability (Honda et al., 2018). Therefore, many recent studies opted to use NWP models for their storm surge simulations (e.g., Rahman et al., 2022; Otaki et al., 2022; Toyoda et al., 2022).
Although the NWP models are preferable over the parametric models, their computational cost is significantly higher than that required for parametric models. Furthermore, an ensemble storm surge prediction from multiple simulations is often desired to provide a range of possible solutions rather than a single predicted value, thus facilitating uncertainty quantification. The ensemble modeling approach will inevitably incur more computational efforts. Consequently, such an obstacle restricts the realization of NWP models in regions with limited computational resources, particularly when the storm surge model is needed for an operational forecasting system. To address the issue, we propose a method based on deep learning to emulate the atmospheric forcings for our storm surge model efficiently.
We utilize typhoon data passing through our study area around Japan from 1981-2012. Based on the International Best Track Archive for Climate Stewardship (IBTrACS) dataset (Knapp et al., 2010;2018), we run the parametric model for the considered typhoon events using the Holland 1980 formula (Holland, 1980). Here, we refer the NWP model to the Japanese 55-year Reanalysis (JRA-55) (Kobayashi et al., 2015) downscaled to a 5-km horizontal resolution named the Dynamical Regional Downscaling Using the JRA-55 Reanalysis (DSJRA-55) (Kayaba et al., 2016) provided by the Japan Meteorological Agency. The reanalysis was created using a state-of-the-art data assimilation method incorporating various observational datasets overlooked in the operational system. Therefore, it yields datasets that better resolve the typhoon’s intensity and track.
We implement a deep learning method based on generative adversarial networks (GAN) (Goodfellow et al., 2014; Isola et al., 2017). The main premise of our deep learning method is to transform the wind and pressure fields by the parametric model into a structure resembling the fields simulated by the NWP model (see Fig. 1). Additionally, we introduce lead-lag parameters to incorporate a forecasting feature in our model. We split the simulated hourly data from the parametric and NWP model pairs into a training set consisting of typhoon events from 1981-2009 (34 events) and a test set comprising the four latest events in the dataset. The resulting atmospheric forcing fields by GAN are then used for the storm surge model based on a regional ocean model known as ROMS (Shchepetkin & McWilliams, 2005). From comparisons with observed water levels during typhoon events on the test set, we achieve comparable accuracy to the NWP-based forcings. The additional computing time to the parametric model by GAN is only approximately one second for a typhoon event using a standard computer.
Although the NWP models are preferable over the parametric models, their computational cost is significantly higher than that required for parametric models. Furthermore, an ensemble storm surge prediction from multiple simulations is often desired to provide a range of possible solutions rather than a single predicted value, thus facilitating uncertainty quantification. The ensemble modeling approach will inevitably incur more computational efforts. Consequently, such an obstacle restricts the realization of NWP models in regions with limited computational resources, particularly when the storm surge model is needed for an operational forecasting system. To address the issue, we propose a method based on deep learning to emulate the atmospheric forcings for our storm surge model efficiently.
We utilize typhoon data passing through our study area around Japan from 1981-2012. Based on the International Best Track Archive for Climate Stewardship (IBTrACS) dataset (Knapp et al., 2010;2018), we run the parametric model for the considered typhoon events using the Holland 1980 formula (Holland, 1980). Here, we refer the NWP model to the Japanese 55-year Reanalysis (JRA-55) (Kobayashi et al., 2015) downscaled to a 5-km horizontal resolution named the Dynamical Regional Downscaling Using the JRA-55 Reanalysis (DSJRA-55) (Kayaba et al., 2016) provided by the Japan Meteorological Agency. The reanalysis was created using a state-of-the-art data assimilation method incorporating various observational datasets overlooked in the operational system. Therefore, it yields datasets that better resolve the typhoon’s intensity and track.
We implement a deep learning method based on generative adversarial networks (GAN) (Goodfellow et al., 2014; Isola et al., 2017). The main premise of our deep learning method is to transform the wind and pressure fields by the parametric model into a structure resembling the fields simulated by the NWP model (see Fig. 1). Additionally, we introduce lead-lag parameters to incorporate a forecasting feature in our model. We split the simulated hourly data from the parametric and NWP model pairs into a training set consisting of typhoon events from 1981-2009 (34 events) and a test set comprising the four latest events in the dataset. The resulting atmospheric forcing fields by GAN are then used for the storm surge model based on a regional ocean model known as ROMS (Shchepetkin & McWilliams, 2005). From comparisons with observed water levels during typhoon events on the test set, we achieve comparable accuracy to the NWP-based forcings. The additional computing time to the parametric model by GAN is only approximately one second for a typhoon event using a standard computer.