Efektivitas Penggabungan Data Lintas Provinsi dan Augmentasi TimeGAN pada Prediksi Harga Beras Berbasis LSTM

Sinta Dwi Lestari (Universitas Teknologi Yogyakarta, Daerah Istimewa Yogyakarta)
Umar Zaky (Universitas Teknologi Yogyakarta, Daerah Istimewa Yogyakarta)

TIN: TERAPAN INFORMATIKA NUSANTARA
Vol 7 No 5 (2026): October 2026

Rice price fluctuations in each province are difficult to predict, although accurate forecasts are needed to anticipate price shocks, while the multivariate data available for several provinces are often short. This study compares two ways of addressing this limitation in a Long Short-Term Memory (LSTM) model: pooling data from three provinces (East Java, Central Java, and West Java) into a single model, and adding synthetic data from a Time-series Generative Adversarial Network (TimeGAN). Monthly data (March 2024 to March 2026) from four sources (weather, rice prices, food reserve stocks, and exchange rates) yielded 25 observations per province. The 84 scenarios under three training conditions (no augmentation, 50%, and 100% augmentation) made 252 experiments, assessed with MAPE, RMSE, and MAE. The pooled model K7 was more accurate than local models in all three provinces. Without augmentation, mean MAPE fell from 2.056% to 1.258%, RMSE from IDR 326.777/kg to IDR 223.018/kg, and MAE from IDR 306.063/kg to IDR 187.795/kg. In contrast, TimeGAN augmentation produced no detectable improvement, and at 100% it tended to raise the pooled models’ MAPE from 1.541% to 1.710%. The synthetic data were easily distinguishable from the real data (mean discriminative score 0.893). For food price monitoring in regions with short data, collecting real data from other provinces appears more promising than building synthetic data, so price-forecasting systems can prioritize pooling data across provinces before adding generative models.