PERFORMANCE OF U-NET AND CONVLSTM FOR 850 HPA WIND FORECASTING IN JAKARTA

Abstract

Very short-range forecasting of 850-hPa winds over tropical urban-coastal regions remains challenging due to complex land-sea circulation, topography, and rapidly changing atmospheric conditions. This study compares a CNN/U-Net baseline and an encoder-decoder Convolutional Long Short-Term Memory (ConvLSTM) model for grid-to-grid prediction of zonal and meridional wind components at a +3-hour horizon over Jakarta and western Java. Experiments use 2025 Global Forecast System outputs in GRIB2 format, with forecast-pair selection, regional extraction, training-only normalization, chronological partitioning, and patch-based sampling. Performance is evaluated using RMSE and MAE, complemented by spatial error maps, strong-wind diagnostics, temporal-gap stress tests, autocorrelation, divergence, and kinetic-energy analyses. ConvLSTM achieves lower RMSE-u (1.018) than U-Net (1.027), while U-Net obtains lower RMSE-v (0.913) than ConvLSTM (0.936). U-Net performs better at 1,144 of 1,681 grid points and degrades more gradually under temporal gaps. Conversely, ConvLSTM better preserves wind peaks, divergence amplitude, and kinetic energy, retaining 93.5% of spectral energy versus 85.4% for U-Net. These results show that model selection should consider spatial, temporal, and physical diagnostics alongside global accuracy metrics.