Peramalan Harga Minyak Sumatran Light Crude Oil Menggunakan Deep Learning “Transformer”

Candrawengi, Ni Luh Putu Ika (2022) Peramalan Harga Minyak Sumatran Light Crude Oil Menggunakan Deep Learning “Transformer”. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6003201012-Master_Thesis.pdf Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

Harga minyak mentah yang terus berubah-ubah baik di dunia maupun di Indonesia tentu berakibat pada kondisi moneter Indonesia. Untuk memproyeksikan harga minyak mentah pada periode berikutnya tentu diperlukan data masa lampau sehingga data harga minya merupakan data time series. Pemodelan pada data time series khususnya pada metode parametrik sangat susah diterapkan kepada data real karena sangat tergantung pada asumsi. Metode deep learning menarik karena fleksibilitasnya untuk tidak membatasi asumsi pada data selain itu juga dapat diterapkan kepada data dengan runtut waktu (time series). Salah satu model dalam deep learning yang dapat digunakan untuk menangkap dependensi jangka panjang adalah model “Transformer”. Model “Transformer” muncul karena adanya kekurangan dari model-model deep learning yang lazim digunakan untuk forecasting seperti RNN, dan LSTM. Model “Transformer” menggunakan algoritma self-attention yang terbagi menjadi dua proses yaitu proses encoding dan proses decoding pada proses training data sehingga fokus hanya pada subset yang paling penting dalam suatu rangkaian panjang yang cocok dan relevan selama pemodelannya. Pada penelitian ini, akan digunakan dua arsitektur “Transformer” yaitu arsitektur “Transformer” Encoder-Decoder dan “Transformer” FC layer yang hanya menggunakan bagian encoder saja. Pada penelitian ini, akan dicobakan beberapa hyperparameter berbeda yaitu dengan menggunakan jumlah head sebesar 16 dan 8, serta percobaan dengan hyperparameter layer berbeda. Hasil penelitian menunjukkan pada data training, model “Transformer” dengan arsitektur encoder FC layer dengan kombinasi head 8 dan layer sebesar 6 memiliki kinerja terbaik RMSE sebesar 31,3253 dan MAPE sebesar 30,8272%. Sedangkan, jika dibandingkan dengan metode LSTM untuk data training¸ diperoleh hasil model LSTM merupakan model yang terbaik untuk memprediksi atau meramalkan harga SLC dengan RMSE sebesar 11,0359 dan MAPE sebesar 16,6611%.
====================================================================================================================================
The price of crude oil that continues to fluctuate both in the world and in Indonesia certainly has an impact on Indonesia's monetary condition. To project crude oil prices in the next period, of course, past data is needed so that the oil price data is time series data. Modeling on time series data, especially on parametric methods, is very difficult to apply to real data because it is very dependent on assumptions. Deep learning methods are attractive because of their flexibility in not limiting assumptions to the data, and they can also be applied to time series data. One model in deep learning that can be used to capture long-term dependencies is the “Transformer” model. The “Transformer” model arises because of the lack of deep learning models commonly used for forecasting such as RNN and LSTM. The "Transformer" model uses a self-attention algorithm that is divided into two processes, namely the encoding process and the decoding process in the training data process so that it focuses only on the most important subsets in a long and relevant series during its modeling. In this research, two "Transformer" architectures will be used, namely the "Transformer" Encoder-Decoder architecture and the "Transformer" FC layer which only uses the encoder part. In this study, several different hyperparameters will be tested, using different numbers of heads for about 8 and 16 and experiments with different layer hyperparameters. The results showed that in training data, the "Transformer" model with an FC layer encoder architecture with a combination of head 8 and layer 6 had the best performance of RMSE of 31.3253 and MAPE of 30.8272%. Meanwhile, when compared with the LSTM method for training data, the results of the LSTM model are the best model for predicting or forecasting SLC prices with RMSE of 11.0359 and MAPE of 16.6611%.

Item Type:	Thesis (Masters)
Additional Information:	RTSt 519.535 Can p-1 2022
Uncontrolled Keywords:	Deep Learning, Self-Attention, Time Series, Transformer
Subjects:	Q Science > QA Mathematics > QA280 Box-Jenkins forecasting
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49101-(S2) Master Thesis
Depositing User:	Mr. Marsudiyana -
Date Deposited:	29 Apr 2026 07:45
Last Modified:	29 Apr 2026 07:45
URI:	http://repository.its.ac.id/id/eprint/132934

Actions (login required)

View Item