A frequency-aware TCN emulator for operational global-to-nearshore wave downscaling driven by correlation-guided forcing

High-resolution nearshore wave forecasting is vital for operations; however, dynamical downscaling with nested wave models is computationally intensive for multisite deployment. While machine learning is used, global-to-nearshore translation is hindered by predictor design: nearest-grid forcing may miss upstream swell connectivity, and use of high-frequency variability may introduce fluctuating components, potentially impairing generalization within a data, decomposition, and model configuration. We introduce a downscaling emulator using frequency awareness and spatial information to map coarse forcing to nearshore significant wave height (Hs). This Hs-focused study evaluates two choices, correlation-guided screening and wavelet-based frequency separation, through controlled ablation studies. Offshore predictors are selected from grid points using correlation-guided screening, and discrete wavelet decomposition preserves low-frequency dynamics and assesses the impact of high-frequency details on downscaling configuration. A Temporal Convolutional Network (TCN) is trained to emulate nearshore targets from nested SWAN simulations. Experiments at Pelabuhan Ratu and Pacitan showed wavelet predictors outperformed raw inputs. Configurations matched targets, with CC = 0.963 and RMSE = 0.03 m at Pelabuhan Ratu, and CC = 0.994 and RMSE = 0.04 m at Pacitan over 14-day test window. In deployment, emulator uses ERA5 or GFS to generate 14-day forecasts without rerunning SWAN.