Spatial-temporal wave height forecast using deep learning and public reanalysis dataset

Abstract

A novel data-driven model is proposed for efficient spatial–temporal significant wave height forecast in West Pacific using deep learning. A multivariate 3-layer LSTM-based architecture following a matrix-to-matrix mapping fashion is applied to establish the underlying spatial–temporal relationship between wind components and significant wave height field for long-term and ocean-scale forecasts. A parallel modeling strategy is employed to reduce the computational expense in which the forecasting domain is flexibly partitioned into finite windows and trained independently as a sub-model. A 30-year hindcast wave dataset released by EMC (Environmental Modeling Center) Operational Wave Models associated with NCEP Climate Forecast System Reanalysis and Reforecast (CFSRR) wind data is used to drive and test the model. A rectangle sea region covering over 9.33 × 106 km2 in West Pacific is regarded as a modeling domain. The data-driven model is validated against a 12-year period historical hindcast data, and global errors are carefully discussed. The forecasting performance of the established model on tropical cyclone wave events is investigated in terms of both spatial pattern and temporal variance of significant wave height. The extreme significant wave height, rough sea occurrence, as well as seasonal and annual characteristics of significant wave height in the West Pacific are carried out using the proposed model and compared with the original hindcast results. The data-driven model shows promising potential to accurately capture the ambiguous patterns and features that are variant in both spatial and temporal dimensions and has obvious superiority over numerical wave modeling in terms of computational efficiency.