Optimization Model for Predicting Stored Grain Temperature Using Deep Learning LSTMs

Abstract

With reliable and accurate grain temperature forecasting models, granary operators could easily make the right decisions to avoid food spoilage. In this study, an analysis of a single hidden layer long short-term memory model, a multi-layer (stacked) long short-term memory model, and its evaluation is presented to determine how accurate it is for forecasting stored grain's temperature from past data. Using temperature sensors, the data is collected over three years in a warehouse in Yunnan, China. There are two datasets: a training dataset and a test dataset. About 40 percent of the data is set aside as a test dataset, while the remaining 60 percent is used as a training dataset. There are several hyper-parameters included in the analysis. By computing the root of the mean square error (RMSE), we can compare the two models. We also use the mean absolute error assessment tool (MAE) and the correlation between predicted and actual values (R2) to evaluate the prediction. In addition to optimizing the number of hidden layers and neurons in each hidden layer, the two models are improved by comparing the actual and predicted models. The experiments we conducted confirm that a single hidden layer can achieve the same or better results than the multilayer (stacked) LSTM when the hyper-parameters are chosen and tuned appropriately, considering the size of the data and the goal.