Using Monte-Carlo Dropout in Deep Neural Networks for Interval Forecasting of Durian Export

Abstract

Interval forecasting is essential because it presents predictions with associated uncertainties, which are not captured by point forecasts alone. In nature, data contain variability due to measurement and random noise. In machine learning, most research focuses on point forecasts, with relatively few studies dedicated to interval forecasting, especially in areas such as agriculture. In this study, durian exports in Thailand are used as a case study. We employed Monte Carlo Dropout (MCDO) for interval forecasting and investigated the impact of various hyperparameters on the performance of Monte Carlo Dropout Neural Networks (MCDO-NNs). Our results were benchmarked against traditional models, such as the Seasonal Autoregressive Integrated Moving Average (SARIMA). The findings reveal that MCDO-NN outperforms SARIMA, achieving a lower root mean squared error of 9,570.24 and a higher R-squared value of 0.4837. The interval forecast width obtained from the MCDO-NN was narrower compared to that of SARIMA. Also, the impact of hyperparameters was observed, and it can serve as guidelines for applying MCDO-NNs to other agricultural datasets or datasets with seasonal and/or trend components.