Penerapan Model Long Short Term Memory (LSTM) untuk Prakiraan Curah Hujan Harian Berbasis Data Automatic Weather Station (AWS) Multi-Stasiun di Kota Medan

Sinaga, Nora Valencia (2026) Penerapan Model Long Short Term Memory (LSTM) untuk Prakiraan Curah Hujan Harian Berbasis Data Automatic Weather Station (AWS) Multi-Stasiun di Kota Medan. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6009232003-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Intensitas hujan yang tinggi kerap memicu kejadian banjir dan genangan di Kota Medan, sehingga curah hujan berperan signifikan dalam mengendalikan dinamika hidrometeorologi di wilayah tropis tersebut. Ketidakpastian pola hujan harian di daerah perkotaan tropis menimbulkan tantangan besar bagi perencanaan tata kota, pengelolaan sumber daya air, serta mitigasi bencana. Sebuah sistem prakiraan hujan yang akurat dan beresolusi tinggi sangat dibutuhkan untuk mendukung pengambilan keputusan cepat dan tepat. Berbasis data Automatic Weather Station (AWS) multi-stasiun di Kota Medan, metode Deep Learning Long Short-Term Memory (LSTM) diterapkan dalam penelitian ini untuk membangun model prakiraan curah hujan harian dengan pendekatan multi-horizon. Data AWS 10 menit periode 2021–2024 dari beberapa stasiun digabungkan dan diselaraskan waktu dalam Universal Unit Coordinate (UTC). Hasil penyelarasan data dalam waktu UTC kemudian dilakukan tahap quality control yaitu pemeriksaan fisik, rate-of-change, konsistensi antarvariabel, dan deteksi spike, serta melakukan imputasi data hilang. Imputasi tersebut diperoleh dari pendekatan hasil interpolasi linier untuk celah pendek dan Multiple Imputation by Chained Equations (MICE) untuk celah panjang. Fitur prediktor dibentuk dari variabel cuaca yaitu suhu, kelembapan, tekanan, angin, radiasi yang diagregasi per jam dan diubah menjadi window waktu input LSTM. Model prakiraan dengan LSTM dirancang dua lapis 128– 64 unit, dropout 0,3, dengan optimizer Adam dilatih untuk memprediksi curah hujan harian hingga 5 hari ke depan. Hasil evaluasi model menunjukkan potensi sistem prakiraan hujan otomatis berbasis data AWS resolusi tinggi untuk mendukung mitigasi bencana hidrometeorologi di wilayah perkotaan tropis. Dengan ambang hujan ≥1 mm/hari, evaluasi kinerja model dilakukan melalui RMSE, MAE, Probability of Detection (POD), False Alarm Ratio (FAR), dan Critical Success Index (CSI); hasilnya mengindikasikan performa yang baik, ditandai RMSE <10 mm/hari pada horizon 1–3 hari, POD >0,80, serta FAR <0,20. Model LSTM ini unggul dibanding model statistik konvensional dengan peningkatan akurasi ~30–40%. Pemanfaatan model LSTM berbasis data AWS multi-stasiun direkomendasikan sebagai bagian dari sistem prakiraan hujan operasional di wilayah perkotaan tropis, dengan pengembangan lebih lanjut untuk meningkatkan akurasi prediksi terutama pada kejadian hujan ekstrem.
=================================================================================================================================
Frequent occurrences of flooding and inundation driven by high rainfall intensity in Medan City underscore the critical role of rainfall in governing hydrometeorological dynamics across tropical regions. Uncertainty in daily rainfall patterns in tropical urban areas poses significant challenges for urban planning, water resource management, and disaster mitigation. An accurate and high-resolution rainfall forecasting system is essential to support rapid and informed decision-making. Based on multi-station Automatic Weather Station (AWS) data from Medan City, Long Short- Term Memory (LSTM) deep learning was employed in this study to construct a multi- horizon daily rainfall forecasting model. Ten-minute AWS data from 2021–2024 from several stations were combined and time-aligned to Universal Unit Coordinates (UTC). The resulting UTC data alignment was then subjected to quality control, including physical inspection, rate-of-change, intervariable consistency, spike detection, and missing data imputation. The input was obtained using a linear interpolation approach for short time slots and Multiple Imputation by Chained Equations (MICE) for long time slots. Predictor features were formed from weather variables—temperature, humidity, pressure, wind, and radiation—aggregated hourly and converted into LSTM input time windows. The LSTM forecast model was designed with two layers of 128– 64 units, with a dropout of 0.3, and the Adam optimizer trained to predict daily rainfall up to five days in advance. Model evaluation results demonstrate the potential of an automated rainfall forecasting system based on high-resolution AWS data to support hydrometeorological disaster mitigation in tropical urban areas. Using a rainfall threshold of ≥1 mm/day, model performance was assessed through Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Probability of Detection (POD), False Alarm Ratio (FAR), and Critical Success Index (CSI), with favorable results characterized by RMSE below 10 mm/day for the 1–3 day horizon, POD exceeding 0.80, and FAR remaining below 0.20. This LSTM model is superior to conventional statistical models with an accuracy increase of ~30–40%. The findings of this study recommend the utilization of an LSTM model based on multi-station AWS data as part of an operational rainfall forecasting system in tropical urban regions, with further development opportunities to enhance prediction accuracy, particularly for extreme rainfall events

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	curah hujan harian, Automatic Weather Station (AWS), Deep learning, LSTM, prakiraan multi-horizon, Kota Medan Daily rainfall, Automatic Weather Station (AWS), deep learning, LSTM,multi-horizon forecasting, Medan City
Subjects:	Q Science > QC Physics > QC866.5 Climatology--Forecasting.
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Physics Engineering > 30101-(S2) Master Thesis
Depositing User:	Nora Valencia Sinaga
Date Deposited:	12 Feb 2026 03:52
Last Modified:	12 Feb 2026 03:52
URI:	http://repository.its.ac.id/id/eprint/132383

Actions (login required)

View Item