Analisis Ketidakpastian Pada Peramalan Curah Hujan Harian di Wilayah Malang Menggunakan LSTM dengan Monte Carlo Dropout

Abimanyu, Ghifari Muammar (2026) Analisis Ketidakpastian Pada Peramalan Curah Hujan Harian di Wilayah Malang Menggunakan LSTM dengan Monte Carlo Dropout. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5003221127_Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Curah hujan harian di wilayah tropis memiliki variabilitas tinggi, banyak nilai nol, dan lonjakan ekstrem sehingga prediksi titik saja belum cukup menggambarkan tingkat keyakinan model. Penelitian ini bertujuan meramalkan curah hujan satu hari ke depan di wilayah Malang menggunakan Long Short-Term Memory (LSTM) berstruktur multi-input single-output serta mengestimasi ketidakpastian epistemik melalui Monte Carlo Dropout (MCD). Data terdiri atas 854 observasi harian BMKG Stasiun Klimatologi Malang periode 5 Januari 2024–7 Mei 2026, dengan masukan berupa riwayat curah hujan, kelembapan relatif rata-rata, lama penyinaran matahari, kecepatan angin rata-rata, suhu minimum, dan suhu maksimum. Pre-processing yang digunakan antara lain, interpolasi linear temporal terbatas, transformasi log1p, dan normalisasi Min-Max. Model menggunakan sliding window delapan hari, sedangkan LSTM-MCD menjalankan 100 inferensi stokastik untuk setiap observasi. Pada 127 data testing, LSTM standar menghasilkan RMSE 11,0045 mm/hari dan MAE 6,5807 mm/hari, sedangkan LSTM-MCD menghasilkan RMSE 10,5324 mm/hari dan MAE 6,2798 mm/hari. Evaluasi ketidakpastian menghasilkan CRPS 5,7043 mm/hari, PICP 0,3228, dan MPIW 6,1745 mm/hari. Hasil tersebut menunjukkan bahwa MCD dapat melengkapi prediksi titik dengan informasi ketidakpastian epistemik, tetapi interval yang dihasilkan masih terlalu sempit dan belum merepresentasikan ketidakpastian prediksi secara keseluruhan.
====================================================================================================================================
Daily tropical rainfall is highly variable, zero-inflated, and prone to extreme spikes; therefore, point forecasts alone cannot represent model confidence. This study aims to forecast one-day-ahead rainfall in Malang using a multi-input single-output Long Short-Term Memory (LSTM) model and estimate epistemic uncertainty through Monte Carlo Dropout (MCD). The data comprised 854 daily observations from the BMKG Malang Climatological Station covering 5 January 2024–7 May 2026. Input variables included rainfall history, mean relative humidity, sunshine duration, mean wind speed, minimum temperature, and maximum temperature. Data preprocessing involved limited temporal linear interpolation, log1p transformation, and Min-Max normalization. The model used an eight-day sliding window, while LSTM-MCD performed 100 stochastic forward passes. On 127 test observations, the standard LSTM produced an RMSE of 11.0045 mm/day and an MAE of 6.5807 mm/day, whereas LSTM-MCD produced an RMSE of 10.5324 mm/day and an MAE of 6.2798 mm/day. Uncertainty evaluation yielded a CRPS of 5.7043 mm/day, a PICP of 0.3228, and an MPIW of 6.1745 mm/day. The results demonstrate that MCD supplements point forecasts with epistemic uncertainty information; however, the resulting epistemic intervals remain too narrow and do not represent total predictive uncertainty.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Curah Hujan Harian, Ketidakpastian Epistemik, Long Short-Term Memory, Monte Carlo Dropout, Peramalan Probabilistik, Daily Rainfall, Epistemic Uncertainty, Long Short-Term Memory, Monte Carlo Dropout, Probabilistic Forecasting
Subjects:	Q Science > QA Mathematics > QA276 Mathematical statistics. Time-series analysis. Failure time data analysis. Survival analysis (Biometry) Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QC Physics > QC925 Rain and rainfall
Divisions:	Faculty of Mathematics, Computation, and Data Science > Statistics > 49201-(S1) Undergraduate Thesis
Depositing User:	Ghifari Muammar Abimanyu
Date Deposited:	05 Aug 2026 08:42
Last Modified:	05 Aug 2026 08:42
URI:	http://repository.its.ac.id/id/eprint/143973

Actions (login required)

View Item