<![CDATA[Banjir rob merupakan ancaman hidrometeorologis yang berdampak signifikan pada degradasi lingkungan pesisir di Kabupaten Brebes. Mitigasi berbasis sistem peringatan dini menuntut adanya prediksi komputasional yang presisi. Namun, pemodelan berbasis machine learning sering kali terhambat oleh keterbatasan ketersediaan data historis (data scarcity). Penelitian ini bertujuan untuk mengembangkan dan membandingkan performa model Deep Learning (LSTM, GRU, BiLSTM) dengan baseline model klasik, yaitu Support Vector Regression (SVR), dalam memprediksi fluktuasi pasang surut multivariat menggunakan 36 bulan data historis (2022–2024). Variabel input terdiri dari kecepatan angin, curah hujan, tekanan barometrik, suhu, dan observasi pasang surut, yang diproses melalui skema pencuplikan temporal (time-windowing) 3 bulan. Evaluasi performa diukur menggunakan RMSE, MAE, R², serta uji signifikansi statistik non-parametrik Wilcoxon Signed-Rank. Hasil eksperimen menunjukkan bahwa SVR secara konsisten mengungguli arsitektur Deep Learning dengan mencatatkan nilai galat terendah (RMSE = 0,1993 m, MAE = 0,1344 m, R² = 0,3434). Keunggulan SVR tervalidasi signifikan secara statistik (p-value = 0,0039 < 0,05). Penelitian ini menyimpulkan bahwa pada kondisi resolusi dataset bulanan yang minor, model regresi klasik yang mengoptimalkan margin toleransi galat beroperasi jauh lebih tangguh dan stabil secara generalisasi dibandingkan model Deep Learning sekuensial yang rentan mengalami overfitting. SVR direkomendasikan sebagai fondasi sistem prediksi tren pasang surut jangka menengah di Kabupaten Brebes. Abstract Tidal flooding (rob) represents a hydrometeorological hazard with significant impacts on coastal environmental degradation in Brebes Regency, Indonesia. Effective mitigation through early warning systems demands reliable computational prediction; however, machine learning-based modeling in this region is frequently constrained by the limited availability of historical records (data scarcity). This study develops and compares the performance of deep learning architectures (LSTM, GRU, BiLSTM) against a classical baseline model, Support Vector Regression (SVR), in predicting multivariate tidal anomaly fluctuations using 36 months of historical data (2022–2024). Input variables comprise wind speed, rainfall, barometric pressure, sea surface temperature, and tidal observations, processed through a 3-month time-windowing scheme. Model performance was evaluated using RMSE, MAE, R2, and the non-parametric Wilcoxon Signed-Rank significance test. Experimental results demonstrate that SVR achieved the lowest prediction error (RMSE = 0.1993 m; MAE = 0.1344 m) among all evaluated models. It is noted, however, that SVR yielded a comparatively lower R2 value (0.3434) relative to several deep learning counterparts, indicating room for further model refinement. The superiority of SVR in error minimization was statistically validated (p = 0.0039 < 0.05). This study concludes that under conditions of limited monthly-resolution datasets, classical regression models grounded in Structural Risk Minimization tend to generalize more stably than sequential deep learning architectures. SVR is recommended as a promising computational baseline for developing a medium-term tidal trend prediction system in Brebes Regency, with validation on higher-resolution data identified as a priority for future research.]]>
Copyrights © 2026
