Effect of Deep Learning on the Prediction of Downward Longwave Radiative Flux in the Arctic

Downward longwave radiative flux (LWD), one of the components determining the surface energy budgets, is important for predicting sea ice properties and climate in the Arctic. In general, large uncertainties in LWD simulations have been found in numerical weather prediction (NWP) models. The LWD was underestimated in the European Centre for Medium-Range Weather Forecasts reanalysis v5 (ERA5), compared to the LWD observations at Hopen in Svalbard, Norway. Based on LWD observations at Hopen, the LWD bias in ERA5 was reduced in the LWD bias of 24 h forecasts in the Polar Weather Research and Forecasting model (PWRF) without and with data assimilation (DA). However, the 24 h LWD forecasts in PWRF still underestimate the LWD observations.

To reduce uncertainties of LWD simulations in the Arctic, a deep learning post-processing model was developed based on the convolutional neural network. The deep learning post-processing model was trained to reduce the LWD bias in ERA5 (2016−2019) with LWD observations at Hopen as truth. The trained deep learning post-processing model was applied to LWD simulations in three independent test data (i.e., ERA5 in 2020, 24 h forecasts in 2020 using PWRF without and with DA). Compared to before deep learning post-processing, the time-averaged root mean square errors of the LWD simulations in the three test data decreased by 17.62%, 14.98%, and 13.14%.

Therefore, deep learning reduces uncertainties in LWD simulations in the Arctic and deep learning can further improve forecasts already using DA. In addition, the deep learning model trained with ERA5 (2016−2019) reduced the LWD biases in 2020 from PWRF without and with DA as well as from ERA5. Therefore, if several NWP models are used to simulate the same atmospheric state, a deep learning model trained using data from one NWP model can be used on data from other NWP models to reduce simulation uncertainties.

Acknowledgments

This study was supported by a National Research Foundation of Korea (NRF) grant funded by the South Korean government (Ministry of Science and ICT) (Grant 2021R1A2C1012572) and Yonsei Signature Research Cluster Program of 2022 (2022-22-0003).