Significant Wave Height Forecasting in the Singapore Strait: Deep Learning Ensemble Stacking with ERA5-IDW Data

Abstract

Accurate significant wave height (SWH) forecasting in the Singapore Strait is critical for maritime safety and marine risk assessment along one of the world’s busiest shipping corridors. This study develops a deep learning ensemble stacking framework for one-step-ahead SWH forecasting using hourly ERA5 reanalysis data spanning January 2021 to December 2025, spatially interpolated to a single target point via Inverse Distance Weighting (IDW) from five ERA5 grid points. Four architectures: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Patch Time Series Transformer (PatchTST), and inverted Transformer (iTransformer) are systematically evaluated against a persistence baseline (RMSE = 0.0188 m), with GRU achieving the best individual performance (RMSE = 0.0153 m, MAPE = 4.58%,  = 0.9814), while ridge regression stacking of GRU with Transformer-based models achieves the overall best performance (RMSE = 0.0151 m, MAPE = 4.54%,  = 0.9820) classified as highly accurate. Transformer-based architectures individually fail to capture SWH variability in this fetch-limited regime yet contribute complementary residual information as stacking base learners. The proposed pipeline establishes a replicable methodology for wave climate forecasting in narrow straits under resolved by global reanalysis products, with potential utility as an environment risk covariate for actuarial modeling in marine cargo insurance underwriting