<![CDATA[Background: Accurate daily rainfall prediction supports flood mitigation in DKI Jakarta, a tropical megacity whose rainfall is shaped by interactions between coastal (sea-breeze) and upstream orographic processes. Gap: Most previous studies rely on single-station data or purely temporal models (e.g., LSTM or GRU), which flatten the spatial layout of observation networks. Method: This study proposes a hybrid architecture that combines a bidirectional convolutional LSTM (Bi-ConvLSTM) with an attention-based GRU (Attention-GRU). Daily data of nine meteorological variables (including lagged rainfall) from four BMKG stations (Kemayoran as the target, Soekarno-Hatta, Tanjung Priok, and Citeko), covering January 2010 to April 2026 (5,964 days), were arranged into a 2×2 grid with a seven-day window, giving an input tensor of shape (7, 2, 2, 9). Results: On a chronologically held-out test set of 1,193 days, the proposed model obtained an RMSE of 16.32 mm, an MAE of 8.20 mm, and an R² of 0.1158, compared with 16.43 mm, 8.45 mm, and 0.1034 for a standard LSTM baseline (RMSE −0.67%, MAE −2.96%); the paired absolute-error difference was significant by the Wilcoxon signed-rank test (p = 0.00155). The gains are modest: ablation variants without attention reached a lower MAE but a higher RMSE, and the model strongly underpredicts rainfall above 20 mm. Contribution: The study provides a reproducible multi-station grid representation for Jakarta and evidence that upstream (Citeko) information contributes most to next-day prediction at Kemayoran. Practical implication: With an inference time of about 18 ms per sample the model is computationally light, but because it underestimates heavy rainfall it is, at this stage, a candidate supporting signal for early warning rather than a stand-alone warning engine.]]>
Copyrights © 2026
