Analisis Clustering dan Peramalan Harga Minyak Goreng Regional dengan Pendekatan Dense-Sparse-Dense Long Short-Term Memory

Abstract

Fluktuasi harga minyak goreng di Indonesia masih terus terjadi, meskipun Indonesia merupakan negara penghasil minyak kelapa sawit terbesar di dunia. Disparitas harga akibat kurang efektifnya sistem distribusi minyak goreng di Indonesia menjadi dasar pentingnya pengelompokan provinsi berdasarkan kemiripan pola pergerakan harga. Pengelompokan provinsi-provinsi dilakukan berdasarkan pola harga minyak goreng menggunakan DBSCAN, K-Means, dan K-Medoids dengan metrik Dynamic Time Warping (DTW). Peramalan harga minyak goreng dilakukan dengan pendekatan Dense-Sparse-Dense Long Short-Term Memory (DSD-LSTM) guna meningkatkan akurasi dan mengurangi overfitting. Data yang digunakan merupakan data harga minyak goreng mingguan dari 34 provinsi Indonesia selama periode Januari 2019 hingga Januari 2025. Data diperoleh dari Pusat Informasi Harga Pangan Strategis Nasional (PIHPS). Hasil evaluasi dari ketiga algoritma clustering menunjukkan bahwa pendekatan K-Means memberikan hasil yang paling optimal, dengan nilai Davies-Bouldin Index (DBI) sebesar 0,54 dan Silhouette score sebesar 39%. Sebanyak 34 provinsi yang dimodelkan menggunakan pendekatan berbasis K-Means menghasilkan nilai MAPE rata-rata hanya sebesar 9,93%, mengungguli pemodelan individu (10,06%). Hasil ini menunjukkan bahwa pendekatan berbasis klaster mampu mempertahankan, bahkan meningkatkan akurasi model. Hasil peramalan 9 Februari – 25 Mei 2025 menunjukkan 14 dari 34 provinsi mengalami kenaikan harga, sementara 20 sisanya mengalami penurunan harga.

The fluctuation of cooking oil prices in Indonesia continues to occur, despite the country being the largest producer of palm oil in the world. Price disparities caused by the inefficiency of the cooking oil distribution system in Indonesia highlight the importance of grouping provinces based on the similarity of their price movement patterns. The clustering of provinces was carried out based on cooking oil price patterns using DBSCAN, K-Means, and K-Medoids with the Dynamic Time Warping (DTW) metric. Price forecasting was performed using the Dense-Sparse-Dense Long Short-Term Memory (DSD-LSTM) approach to improve accuracy and reduce overfitting. The data used consists of weekly cooking oil prices from 34 provinces in Indonesia covering the period from January 2019 to January 2025. The data was obtained from the National Strategic Food Price Information Center (PIHPS). Evaluation results from the three clustering algorithms show that the K-Means approach yielded the most optimal results, with a Davies-Bouldin Index (DBI) of 0.54 and a Silhouette score of 39%. A total of 34 provinces modeled using the K-Means-based approach achieved an average MAPE of only 9.93%, outperforming the individual modeling approach (10.06%). These results indicate that the cluster-based approach can maintain, and even improve, model accuracy. The forecasting results for the period from February 9 to May 25, 2025, show that 14 out of 34 provinces experienced price increases, while the remaining 20 experienced price decreases.