Transfer learning applied to tidal current predictions with limited data availability

Accurate prediction of tidal current velocities is crucial for various applications, such as estimating energy production in tidal power projects. However, Acoustic Doppler current profiler (ADCP) campaigns are often expensive and can lead to incomplete data. This study addresses that limitation by transferring knowledge between estuaries: a long short-term memory (LSTM) network is pre-trained on a data-rich site and fine-tuned on a data-scarce target, using only astronomically predictable inputs, namely decimal hour, tidal height, a harmonic velocity prediction (HVP), and lunar phase. On held-out data, using a strictly temporal validation protocol, the transferred model reaches R 2 = 0.932 and RMSE = 0.140 m/s under five-fold walk-forward validation, reducing RMSE by 48% with respect to an identical network trained from scratch and by 35% with respect to a four-constituent harmonic-analysis baseline, and confining the spread across random seeds to a narrow band instead of the wide dispersion seen without transfer. The unmeasured portion of the lunar cycle is then reconstructed with a calibrated prediction band. An ablation shows that the harmonic input becomes increasingly valuable as the fine-tuning record shrinks, compensating for data scarcity.