VAE Based Non-Autoregressive Transformer Model for Sea Ice Concentration Forecast

• Main Authors: Wu, Da, Lang, Xiao, Mao, Wengang, Zhang, Di, Zhang, Jinfen, Liu, Rong
• Language: unknown
• Published: 2022
• Subjects: Marine Engineering; Probability Theory and Statistics; Other Electrical Engineering; Electronic Engineering; Information Engineering; Variational Autoencoder; Arctic shipping; Transformer; Non autoregressive model; Sea ice concentration; Time series analysis; Arctic; Vae; Northern Sea Route; Sea ice
• Online Access: https://research.chalmers.se/en/publication/fec6ae7c-74c4-4d38-beb9-184e99a99eb6

Arctic shipping along the Northern Sea Route (NSR) relies on reliable and precise sea ice concentration (SIC) information. However, the SIC forecast used to assist NSR shipping is often inaccurate. This study proposes a VAE based Non-Autoregressive Transformer model for the long-term SIC forecast task. The proposed model ensembles the deep generative model of Variational Autoencoder and the deep learning model of Transformer to overcome the temporal delay and accumulative error problems shown in traditional time series models. Model validation has been conducted to compare the forecasts of SIC with other machine learning and deep learning models. The proposed model outperforms the compared models in terms of different metrics. The proposed VAE based Non-Autoregressive Transformer can be used for long-term SIC forecast and achieve stable and good accuracy predictions.