Comparative evaluation of statistical and deep learning methods for high-frequency radar surface current forecasting in a narrow tropical strait

Abstract

Accurate prediction of ocean surface currents is essential for maritime safety in narrow waterways with strong tidal forcing. This study presents a three-phase evaluation of forecasting methods for high-frequency (HF) radar surface currents in the Sunda Strait, Indonesia, where tidal currents regularly exceed 150 cm s −1 . Phase 1 compares twelve methods – persistence, tidal harmonics, classical time series, a reduced-rank spatio-temporal statistical model, shallow machine learning, and deep learning – for one-step-ahead prediction; evaluated consistently over all 291 grid cells and the full test period, the deep learning models achieve the lowest root-mean-square errors (RMSE) for both components, with the convolutional neural network (CNN) reaching 11.31 cm s −1 for the zonal component (skill score 0.50 relative to persistence) and the convolutional neural network–gated recurrent unit (CNN-GRU) reaching 15.44 cm s −1 for the meridional component (skill score 0.42); the pointwise classical statistical models (autoregressive integrated moving average, ARIMA; exponential smoothing) fall well below, while an empirical-orthogonal-function vector autoregression (EOF-VAR) that models spatial correlation closes most of the gap to deep learning, indicating that the neglect of spatial structure – not of nonlinearity – is the principal limitation of the classical baselines. Phase 2 examines lookback window length ( T  = 3, 6, 12 h); CNN-gated recurrent unit (CNN-GRU) improves monotonically with longer lookback, reaching 11.21 and 15.41 cm s −1 at T  = 12, while standalone CNN and GRU show no benefit. Phase 3 extends prediction to six-hour nowcasting using direct multi-step and autoregressive (ConvLSTM-ED, BiEF) architectures; under a controlled comparison (five seeds, identical training budget on one GPU), accuracy is governed by model capacity and plateaus near 1 M parameters (the 1.00 M CNN-GRU-MS-Small reaching 18.39 and 22.07 cm s −1 ; skill scores 0.77 and 0.72), and at matched capacity the direct and autoregressive architectures are statistically indistinguishable. The direct multi-step model is nonetheless preferred for operational nowcasting because it trains ∼2.5 times faster and runs ∼4.5 times faster at inference (a single forward pass versus sequential decoding). Diurnal error analysis, corroborated by concurrent wind observations from three coastal weather stations, reveals that zonal prediction errors correlate positively with afternoon sea-breeze wind enhancement (Spearman r s  = 0.48 across the 24 diurnal-mean hours), while meridional errors follow a distinct diurnal pattern not explained by wind forcing. Relative errors of 3 %–7 % of the observed speed range are of the same order as those reported in calmer environments – though this normalisation partly reflects the strait's very large speed range – suggesting that deep learning does not break down in energetic strait settings.