Performa Peramalan Curah Hujan Harian di Bali Menggunakan Model Hibrida Prophet-XGBoost dengan Koreksi Residual: Studi Komparatif dengan Model Tunggal

Ghaitsa, Khansa Tsabita (2026) Performa Peramalan Curah Hujan Harian di Bali Menggunakan Model Hibrida Prophet-XGBoost dengan Koreksi Residual: Studi Komparatif dengan Model Tunggal. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5006221094-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Curah hujan harian memiliki karakteristik yang fluktuatif, tidak simetris, serta mengandung kejadian ekstrem yang relatif jarang terjadi. Kondisi tersebut menyebabkan peramalan curah hujan harian menjadi tantangan karena model tidak hanya perlu menangkap pola musiman, tetapi juga perlu merespons variasi harian yang bersifat tidak linear. Penelitian ini bertujuan untuk membandingkan performa model Prophet, XGBoost, dan hybrid Prophet–XGBoost dengan pendekatan koreksi residual dalam meramalkan curah hujan harian pada titik representatif Bali Barat. Data yang digunakan merupakan data curah hujan harian dari NASA POWER periode 1 Januari 2015 hingga 31 Desember 2024. Data dibagi secara kronologis menjadi data latih dan data uji dengan rasio 80:20. Tahap preprocessing dilakukan melalui pemeriksaan kelengkapan tanggal, duplikasi, missing value, dan nilai tidak valid. Selanjutnya, dilakukan feature engineering untuk membentuk fitur kalender dan musiman, fitur lag, rollingwindow, serta indikator historis hujan dan hujan lebat. Model Prophet digunakan untuk menangkap pola tren dan musiman, XGBoost digunakan untuk memodelkan hubungan nonlinier berbasis fitur historis, sedangkan model hybrid menggunakan XGBoost untuk memodelkan residual dari hasil prediksi Prophet. Evaluasi model dilakukan menggunakan MAE, RMSE, R², dan SMAPE. Hasil penelitian menunjukkan bahwa Prophet Tunggal menghasilkan MAE sebesar 5,86 mm/hari, RMSE sebesar 14,79 mm/hari, dan R² sebesar 0,15pada data uji. Model XGBoost tunggal menghasilkan MAE sebesar 4,58 mm/hari, RMSEsebesar 14,60 mm/hari, dan R² sebesar 0,17. Sementara itu, model hybrid Prophet–XGBoost residual menghasilkan MAE sebesar 5,22 mm/hari, RMSE sebesar 14,62 mm/hari, dan R²sebesar 0,16. Berdasarkan hasil evaluasi, XGBoost tunggal dengan lima fitur utama dan transformasi log1p dipilih sebagai model terbaik karena menghasilkan rata-rata kesalahan prediksi terendah pada data uji. Hasil forecasting menggunakan model terbaik untuk periode 1–30 Januari 2025 menghasilkan prediksi curah hujan harian pada rentang 5,47–9,63 mm/hari, dengan rata-rata 8,62 mm/hari. Hasil ini menunjukkan bahwa XGBoost mampu memberikan performa yang lebih baik dibandingkan Prophet dan hybrid residual dalam memodelkan curah hujan harian pada titik representatif Bali Barat, meskipun seluruh model masih memiliki keterbatasan dalam memprediksi lonjakan curah hujan ekstrem.
=====================================================================================================================================
Daily rainfall is characterized by high fluctuation, asymmetrical distribution, and relatively rare extreme events. These characteristics make daily rainfall forecasting challenging because the model must not only capture seasonal patterns but also respond to nonlinear daily variations. This study aims to compare the performance of Prophet, XGBoost, and a Prophet–XGBoost hybrid model with a residual correction approach in forecasting daily rainfall at a representative point in West Bali. The data used in this study are daily rainfall data from NASAPOWER for the period from January 1, 2015, to December 31, 2024. The data were chronologically divided into training and testing sets with an 80:20 ratio. The preprocessing stage included checking date completeness, duplicates, missing values, and invalid values. Furthermore, feature engineering was conducted to construct calendar and seasonal features, lag features, rolling window features, and historical indicators of rainfall and heavy rainfall. Prophet was used to capture trend and seasonal patterns, XGBoost was used to model nonlinear relationships based on historical features, while the hybrid model used XGBoost to model the residuals from Prophet predictions. Model evaluation was conducted using MAE, RMSE, R²,and SMAPE. The results show that the single Prophet model produced an MAE of 5.86 mm/day,an RMSE of 14.79 mm/day, and an R² of 0.15 on the testing data. The single XGBoost model produced an MAE of 4.58 mm/day, an RMSE of 14.60 mm/day, and an R² of 0.17. Meanwhile, the Prophet–XGBoost residual hybrid model produced an MAE of 5.22 mm/day, an RMSE of14.62 mm/day, and an R² of 0.16. Based on the evaluation results, the single XGBoost model with five main features and log1p transformation was selected as the best model because it produced the lowest average prediction error on the testing data. Forecasting results using the best model for the period from January 1 to January 30, 2025, produced daily rainfall predictions ranging from 5.47 to 9.63 mm/day, with an average of 8.62 mm/day. These results indicate that XGBoost was able to provide better performance than Prophet and the residual hybrid model in modeling daily rainfall at the representative point in West Bali, although all models still have limitations in predicting extreme rainfall spikes.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Peramalan Curah Hujan Harian, Hybrid Prophet-XGBoost, Prophet, Time series cross-validation, XGBoost, Daily Rainfall, Hybrid Prophet-XGBoost, Forecasting, Prophet, Time series cross-validation, XGBoost.
Subjects:	H Social Sciences > HA Statistics > HA30.3 Time-series analysis Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QA Mathematics > QA280 Box-Jenkins forecasting Q Science > QC Physics > QC866.5 Climatology--Forecasting.
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Actuaria > 94203-(S1) Undergraduate Thesis
Depositing User:	Khansa Tsabita Ghaitsa
Date Deposited:	17 Jul 2026 04:36
Last Modified:	17 Jul 2026 04:53
URI:	http://repository.its.ac.id/id/eprint/135274

Actions (login required)

View Item