Peramalan Curah Hujan menggunakan Transformer berbasis Transfer Learning dari Data Regional pada Wilayah dengan Data Terbatas di Kabupaten Muna Barat

Epifania, Ni Luh Nafa Epifania (2026) Peramalan Curah Hujan menggunakan Transformer berbasis Transfer Learning dari Data Regional pada Wilayah dengan Data Terbatas di Kabupaten Muna Barat. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5006221017-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Curah hujan merupakan salah satu unsur iklim yang berperan penting dalam sektor pertanian, pengelolaan sumber daya air, dan ketahanan pangan, khususnya di wilayah tropis seperti Indonesia. Peramalan curah hujan pada skala wilayah administratif kecil masih menghadapi kendala utama berupa keterbatasan data historis, terutama pada wilayah yang memiliki rekaman pengamatan pendek dan tidak kontinu. Kondisi ini menyulitkan penerapan model peramalan berbasis deep learning yang umumnya membutuhkan data dalam jumlah besar. Penelitian ini menerapkan model Transformer dengan pendekatan Transfer Learning untuk mengatasi keterbatasan tersebut. Data curah hujan regional Pulau Muna periode 1981–2025 yang bersumber dari NASA digunakan sebagai domain sumber pada tahap Pre-training, sedangkan data curah hujan Kabupaten Muna Barat periode 2015–2025 dari BPS digunakan sebagai domain target pada tahap fine-tuning. Untuk mengevaluasi efektivitas pendekatan ini, dilakukan perbandingan antara model Transformer dengan Transfer Learning dan model Transformer tanpa Transfer Learning menggunakan metrik Nash-Sutcliffe Efficiency (NSE), Root Mean Square Error (RMSE), dan symmetric Mean Absolute Percentage Error (sMAPE). Hasil evaluasi menunjukkan bahwa model Transformer tanpa Transfer Learning menghasilkan nilai NSE sebesar 0,5494, sMAPE sebesar 46,2870%, dan RMSE sebesar 83,3595 mm. Sementara itu, model Transformer dengan Transfer Learning menghasilkan nilai NSE sebesar 0,6197 dan RMSE sebesar 80,4596 mm yang lebih baik, meskipun nilai sMAPE meningkat menjadi 70,7079% akibat sensitivitas metrik terhadap nilai curah hujan rendah. Model terbaik selanjutnya digunakan untuk melakukan peramalan curah hujan periode Januari 2026 hingga Desember 2027. Hasil peramalan menunjukkan bahwa model dengan Transfer Learning mampu menghasilkan pola musiman yang lebih konsisten dan realistis dibandingkan model tanpa Transfer Learning. Penelitian ini memberikan bukti empiris bahwa pendekatan Transfer Learning berbasis data satelit berpotensi mengatasi keterbatasan data lokal dalam peramalan curah hujan pada skala wilayah terbatas.
===================================================================================================================================
Rainfall is a critical climate variable influencing agriculture, water resource management, and food security, particularly in tropical regions such as Indonesia. Forecasting rainfall at small administrative scales remains challenging due to limited historical data, especially in areas with short and discontinuous observation records. This condition hinders the application of deep learning-based forecasting models, which generally require large amounts of Training data.
This study applies a Transformer model with a Transfer Learning approach to address this limitation. Regional rainfall data from Muna Island covering the period 1981–2025, sourced from NASA, was used as the source domain for Pre-Training, while rainfall data from West Muna Regency covering 2015–2025 from BPS was used as the target domain for Fine-tuning. To evaluate the effectiveness of this approach, a comparison was conducted between the Transformer model with Transfer Learning and the Transformer model without Transfer Learning, using Nash-Sutcliffe Efficiency (NSE), Root Mean Square Error (RMSE), and symmetric Mean Absolute Percentage Error (sMAPE) as evaluation metrics. The results show that the Transformer without Transfer Learning achieved an NSE of 0,4538, sMAPE of 65,1894%, and RMSE of 96,4212 mm. The Transformer with Transfer Learning produced a higher NSE of 0.6197 and a lower RMSE of 80.4596 mm, although sMAPE increased to
70.7079% due to the metric's sensitivity to near-zero rainfall values. The best-performing model was subsequently used to forecast monthly rainfall from January 2026 to December 2027. Forecasting results indicate that the Transfer Learning model produced more consistent and realistic seasonal patterns compared to the model trained without transfer knowledge. This study provides empirical evidence that satellite-based Transfer Learning offers a viable approach to overcoming local data limitations in rainfall forecasting for data-scarce regions.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Curah Hujan, Kabupaten Muna Barat, Peramalan, Transfer Learning, Transformer. Forecasting, Rainfall, Transfer Learning, Transformer, West Muna Regency.
Subjects:	G Geography. Anthropology. Recreation > G Geography (General) > G70.217 Geospatial data Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Actuaria > 94203-(S1) Undergraduate Thesis
Depositing User:	Ni Luh Nafa Epifania
Date Deposited:	23 Jul 2026 08:21
Last Modified:	23 Jul 2026 08:21
URI:	http://repository.its.ac.id/id/eprint/135367

Actions (login required)

View Item