Pengembangan Model Long-Term Time Series Forecasting Berbasis Frequency-Domain Attention In Two Horizons

Wahyudi, Bimantara Tito (2025) Pengembangan Model Long-Term Time Series Forecasting Berbasis Frequency-Domain Attention In Two Horizons. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025201227-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Dalam dunia long-term time series forecasting, berbagai model telah dikembangkan untuk mencapai performa prediksi yang optimal. Akan tetapi, tantangan dalam data seperti noise, fluktuasi acak, serta pola seasonal yang tidak teratur dan dinamis sering kali menghambat akurasi model, terutama dalam rentang prediksi yang panjang. Penelitian ini mengusulkan pengembangan model long-term time series forecasting berbasis Frequency-Domain Attention in Two Horizons (FAITH) dengan melakukan modifikasi arsitektur pada dua komponen utama. Pertama, bagian dekomposisi diubah menggunakan EMA decomposition untuk memisahkan komponen trend dan seasonal secara lebih efektif. Kedua, jalur pemrosesan trend dimodifikasi menggunakan MLP-based trend stream untuk menangkap long-term temporal. Kedua komponen ini diadopsi dari model xPatch dan digabungkan dalam satu arsitektur baru yang disebut sebagai Improved FAITH. Evaluasi model dilakukan melalui serangkaian uji coba pada 8 dataset, 4 prediction length, dan 2 metrik evaluasi. Hasil uji coba menunjukkan bahwa Improved FAITH berhasil mengungguli FAITH versi original dalam 48 dari 64 skenario dengan rata-rata peningkatan performa sebesar 1,003%. Selain itu, model ini juga menunjukkan peningkatan performa dibandingkan dengan lima model long-term forecasting terkemuka, yaitu Reformer (48,096%), Informer (37,325%), Autoformer (24,070%), FEDformer (12,238%), dan LTSF-Linear (1,267%). Hasil uji coba ini menunjukkan bahwa Improved FAITH merupakan model yang kompetitif dan mampu mengatasi berbagai tantangan utama dalam long-term time series forecasting.
=================================================================================================================================
In the field of long-term time series forecasting, various models have been developed to achieve optimal predictive performance. However, data-related challenges such as noise, random fluctuations, and irregular or dynamic seasonal patterns often disturb model accuracy, especially over extended prediction horizons. This study proposes the development of a long-term forecasting model based on Frequency-Domain Attention in Two Horizons (FAITH) by modifying two key architectural components. First, the decomposition module is replaced with EMA decomposition to more effectively separate the trend and seasonal components. Second, the trend processing path is modified using an MLP-based trend stream to better capture long-term temporal dependencies. Both components are adopted from the xPatch model and combined into a new architecture referred to as Improved FAITH. The model was evaluated through a series of experiments involving 8 datasets, 4 prediction lengths, and 2 evaluation metrics. The results show that Improved FAITH outperformed the original FAITH model in 48 out of 64 scenarios, with an average performance improvement of 1.003%. Furthermore, the model also demonstrated performance increase compared to five leading long-term forecasting models, Reformer (48.096%), Informer (37.325%), Autoformer (24.070%), FEDformer (12.238%), and LTSF-Linear (1.267%). These results indicate that Improved FAITH is a competitive model capable of addressing key challenges in long-term time series forecasting.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Prediksi deret waktu, domain frekuensi, Transformer, FAITH, Time series forecasting, frequency domain, Transformer, FAITH
Subjects:	T Technology > T Technology (General) T Technology > T Technology (General) > T174 Technological forecasting
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Bimantara Tito Wahyudi
Date Deposited:	24 Jul 2025 09:15
Last Modified:	27 Jul 2025 14:02
URI:	http://repository.its.ac.id/id/eprint/121342

Actions (login required)

View Item