Abstract
Food storage security is critical to the national economy and people's lives. The environmental parameters of a granary should be accurately monitored in order to provide a better preservation environment for food storage. In this paper, we use temperature sensors to measure and collect grain temperature data for a period of 423 days from a real world granary, and collect the corresponding meteorological data from China Meteorological Data Network. We propose to leverage meteorological data to predict the average temperature of the grain pile with machine learning algorithms: a support vector regression (SVR) approach and an adaptive boosting (AdaBoost) approach. We incorporate different kernel functions in the SVR model and choose the appropriate base-estimator and the number of estimators in the AdaBoost model. We then analyze the correlation between a large amount of historical data from the granary and the corresponding meteorological forecast data based on the Pearson correlation coefficient. We find that there are strong correlations between some meteorological factors. In order to eliminate redundant information, we reduce the dimension of data by principal components analysis (PCA), and compare the prediction models before and after PCA dimension reduction. The results show that the proposed approaches can achieve a high accuracy and the Adboost method after PCA dimension reduction achieves the best performance.