Forecasting of garlic price based on DA-RNN using attention weight of temporal fusion transformers

Abstract

Garlic is a major condiment vegetable grown in South Korea. The price of garlic has a great impact on Korean society and the economy, which requires price stabilization through preemptive supply and demand management. Therefore, the government attempts to keep the price adjusted according to the predicted production cost. However, classic statistical models or well-known deep learning models have lower forecast accuracy when the number of input factors increases. The aforementioned issue could make analysis approaches and their implementation difficult, and the government would confront failure in proper supply and demand management. To solve this problem, we propose a new hybrid deep-learning approach that employs well-known attention models. Recent attention models have achieved outstanding performance in time-series dataset forecasting. However, when input datasets contain dozens or hundreds of variables, the forecasting performance cannot be guaranteed because the prediction accuracy decreases. In this study, a novel approach utilizing attention weights for forecasting prices is introduced. Experience shows that forecasting accuracy can be improved using the proposed model, which deals with different variables related to garlic prices, such as atmospheric conditions, logistics processes, and environmental circumstances. The proposed approach and its model contribute to forecasting outputs for different research domains by using a variety of attention weight models.