Pacific Decadal Oscillation Forecasting With Spatiotemporal Embedding Network

The Pacific decadal oscillation (PDO) is a decadal variability phenomenon occurring in the North Pacific Ocean. It has substantial impacts on marine ecosystems and the global climate. Due to the high complexity and unclear evolution mechanism, the accurate long-term prediction of PDO remains a challenge. In this paper, a deep spatiotemporal embedding network (DSEN) is proposed to extract the spatiotemporal features from historical climate data and achieve end-to-end forecasting of the PDO index. The spatiotemporal features are recursive in the continuous forecasting of the PDO index on seasonal time scales, thus the cumulative error is largely reduced. During the test period of 39 years (1982-2020), our model can skillfully predict the PDO index up to 1 year, outperforming six methods used as benchmark. By contrast with physically-based methods, DSEN can accurately predict the PDO index from a data-driven perspective.