Perbandingan Model LSTM dan 1D-CNN dalam Peramalan Harga Beras Bulanan di Indonesia

Gonta, Gladys Akiko (2026) Perbandingan Model LSTM dan 1D-CNN dalam Peramalan Harga Beras Bulanan di Indonesia. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5003221016-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Harga beras merupakan salah satu indikator ketahanan pangan nasional yang perlu dipantau untuk memahami pergerakan harga dari waktu ke waktu. Penelitian ini membandingkan dua arsitektur deep learning, yaitu Long Short-Term Memory (LSTM) dan One-Dimensional Convolutional Neural Network (1D-CNN), dalam meramalkan harga beras bulanan di tingkat perdagangan besar (grosir) Indonesia. Data yang digunakan merupakan deret waktu univariat periode Januari 2010 hingga Desember 2025 sebanyak 192 observasi dengan satuan rupiah per kilogram. Setiap arsitektur diuji menggunakan dua strategi peramalan multilangkah, yaitu recursive dan direct Multiple-Input Multiple-Output (MIMO), tiga ukuran window masukan (12, 24, dan 36 bulan), serta dua bentuk data, yaitu harga asli (level) dan hasil transformasi differencing orde satu, sehingga diperoleh 24 konfigurasi yang dibandingkan. Setiap konfigurasi dijalankan sebanyak sepuluh kali dengan seed yang berbeda dan diringkas menggunakan nilai rata-rata beserta confidence interval 95%. Kinerja model diukur menggunakan Root Mean Squared Error (RMSE) dan Mean Absolute Percentage Error (MAPE) pada skala harga asli. Hasil penelitian menunjukkan bahwa pemodelan langsung pada data level belum mampu mengikuti tren kenaikan harga aktual, sedangkan transformasi differencing menghasilkan kinerja yang lebih baik pada seluruh konfigurasi model yang diuji. Konfigurasi terbaik diperoleh pada LSTM dengan strategi recursive dan window 12 bulan menggunakan data hasil transformasi differencing, dengan RMSE sebesar 821,1 ± 11,1 dan MAPE sebesar 4,55% ± 0,05%. Pada data hasil transformasi differencing, model LSTM secara konsisten menghasilkan kinerja yang lebih baik dibandingkan 1D-CNN, sedangkan strategi peramalan yang memberikan hasil terbaik berbeda pada setiap arsitektur, yaitu recursive untuk LSTM dan direct MIMO untuk 1D-CNN. Model terbaik selanjutnya digunakan untuk meramalkan harga beras periode Januari hingga Desember 2026 dan menghasilkan prediksi kenaikan harga secara bertahap dari Rp14.177 pada Januari 2026 menjadi Rp14.409 pada Desember 2026 atau meningkat sebesar 1,75% dibandingkan harga aktual Desember 2025.
=======================================================================================================================================
Rice price is one of the indicators of national food security that needs to be monitored to understand price movements over time. This study compares two deep learning architectures, namely Long Short-Term Memory (LSTM) and One-Dimensional Convolutional Neural Network (1D-CNN), for forecasting monthly rice prices at the wholesale level in Indonesia. The data used are univariate time series data from January 2010 to December 2025, consisting of 192 observations measured in Indonesian Rupiah per kilogram. Each architecture was evaluated using two multi-step forecasting strategies, namely recursive and direct Multiple-Input Multiple-Output (MIMO), three input window sizes (12, 24, and 36 months), and two data forms, namely original price data (level) and first-order differenced data, resulting in 24 configurations for comparison. Each configuration was executed ten times using different seeds and summarized using the average value along with a 95% confidence interval. Model performance was evaluated using Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE) on the original price scale. The results showed that models trained directly on level data were not yet able to follow the increasing trend of actual prices, while first-order differencing produced better performance across all evaluated configurations. The best configuration was obtained from LSTM with the recursive strategy and a 12-month input window using differenced data, achieving an RMSE of 821.1 ± 11.1 and a MAPE of 4.55% ± 0.05%. On differenced data, LSTM consistently achieved better performance than 1D-CNN, while the most effective forecasting strategy varied depending on the architecture, with recursive performing best for LSTM and direct MIMO performing best for 1D-CNN. The best model was then used to forecast rice prices for the January to December 2026 period, producing a gradual increase from Rp14,177 in January 2026 to Rp14,409 in December 2026, or an increase of 1.75% compared with the actual price in December 2025.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	LSTM, 1D-CNN, Differencing, Harga Beras, Peramalan, Rice Price, Forecasting
Subjects:	Q Science > QA Mathematics > QA276 Mathematical statistics. Time-series analysis. Failure time data analysis. Survival analysis (Biometry)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49201-(S1) Undergraduate Thesis
Depositing User:	Gladys Akiko Gonta
Date Deposited:	04 Sep 2026 02:09
Last Modified:	04 Sep 2026 02:09
URI:	http://repository.its.ac.id/id/eprint/142906

Actions (login required)

View Item