Grain Temperature Prediction Based on GRU Deep Fusion Model

Abstract

Temperature is an essential quality index in storage. Prediction of temperature can help the grain storage industry to apply the appropriate operations such as ventilation or drying to improve the quality of grain and extend the suitable storage time. Traditional machine learning methods usually cannot accurately predict the temperature data of the grain considering the complexity of environmental factors and grain warehouse conditions. To make better use of the temporal data such as temperature/humidity information of grain itself and its environment, this paper proposes a gated recurrent unit (GRU)-based algorithm to predict the change of the data. The grain warehouse environmental data are collected by multi-functional sensors inside a grain depot, including temperature, humidity, wind speed, air pressure, etc. Some of these data features such as rain or snow days are sparse data features. Excessive sparse features can affect the training accuracy of the model. At the same time, due to sensor aging or extreme weather conditions, the data collected may not be accurate, and the data contain noise, which also has a significant impact on the training of the model. To improve the performance of the proposed GRU framework, multivariate linear regression is used for feature generation to optimize the volatility of weather data, strengthen and construct the characteristics of datasets, and wavelet filtering is used to denoise the corresponding features. This paper focuses on the data sparse and noise problem and applies the MLR and wavelet filtering to improve the GRU prediction framework for grain warehouse temporal data. According to our experiment, the temperature prediction results based on the GRU deep fusion model have better improvement in prediction accuracy and time than the existing neural network algorithms such as long–short-term memory (LSTM), GRU, and transformer.