冷链可以有效维持易腐食品品质、保障质量安全、降低过程损耗，冷链环境监控对于充分发挥冷链物流效能至关重要。当前，冷链环境监控技术可以满足多点、无线、实时等监测需求，并通过与新一代智能信息技术结合开始向冷链物流环境精准预测方向快速发展。本文针对冷链运输温度预测问题，从挖掘历史数据时序信息角度出发，提出了一种基于门控循环单元网络 （Gated recurrent unit，GRU） 的冷链运输温度时序预测方法。首先，滤除冷链运输温度时序数据的异常值和缺失值，利用拉格朗日插值法进行填补修正后归一化处理；然后，根据三种不同时间序列预测尺度的GRU神经网络模型预测性能对比结果，选择构建GRU时序预测模型用于预测冷链中实验冷藏厢体未来10 min的温度变化，并与循环神经网络 （Recurrent neural network，RNN） 模型、反向传播神经网络 （Back propagation，BP） 模型进行预测准确性对比试验。对比冷链厢体温度真实值与预测值发现，基于 GRU 神经网络模型的对应厢体预测温度均方根误差和平均绝对误差分别为 0.156 和 0.760 ℃，平均绝对百分比误差为0.236%，与其他模型相比，以上误差指标值均处于相对最低水平；在温度时间序列预测模型中，GRU时序预测模型表现出更优的预测效果。该研究成果对于食品冷链物流中预测预警食品安全、精细控制冷链环境具有重要的实际指导意义。

Cold-chain transportation maintains the quality and safety of perishable food, thus reducing process losses. Monitoring the cold-chain environment is very important in ensuring its optimal performance. At present, the technology for monitoring the cold-chain environment enables multi-point, wireless, real-time measurements. Combined with new-generation intelligent information technology, there have been developments toward accurate predictions of the cold-chain environment. With a focus on predicting the cold-chain transportation temperature, this study proposes a method of time series temperature prediction based on a gated recurrent unit (GRU) artificial neural network that mines the time series information of historical data. First, the outliers and missing values of temperature data collected from refrigerated compartments are filtered out, then filled and corrected by Lagrange interpolation before being normalized. Comparisons of the prediction performance of the GRU models using time series data on three different scales lead to the construction of a GRU model for predicting the temperature changes of an experimental refrigerated compartment over the next 10 min. Finally, a recurrent neural network(RNN) and backpropagation neural network(BP) are used for comparative experiments. Comparing the actual temperatures with the predicted values, we find that the root mean square and average absolute errors of the compartment temperature predictions using the GRU model are 0.156°C and 0.760℃, respectively, and the average absolute percentage error is 0.236%. The GRU model exhibits better prediction performance than the recurrent neural network and backpropagation neural network. The research results are of significance for accurate early warning of food safety and precise control of the cold-chain environment.