基于TiDE-PatchTST模型的柑橘冷藏效率时序预测模型优化

Optimization of Citrus Cold Storage Efficiency Time-series Prediction Model Based on TiDE-PatchTST

摘要:

柑橘低温贮藏过程中果实温度波动是引发果品品质安全风险与增加制冷能耗的关键因素，同时果品品质与制冷能耗也是评判柑橘冷藏效率的重要评价指标，实现两者动态预测可为科学预知与精准优化柑橘冷藏效率提供可靠支持。本文提出一种基于PatchTST的柑橘冷藏效率时序预测模型。首先，基于自注意力机制和独立预测方法(Channel independent,CI)构建基础PatchTST模型；其次，通过融合基础PatchTST模型与TiDE模型中的协变量特征提取模块，实现对多元时序数据集中全部序列的特征提取，并有效改进模型预测精度；最后，基于皮尔森相关性分析方法定量分析冷库制冷参数与能耗、柑橘温度的相关性，确定TiDE-PatchTST模型输入参数，并基于5000组实验数据实现多种模型训练与测试，对比验证TiDE-PatchTST模型的准确性与优越性。结果表明，基于TiDE-PatchTST模型的冷库能耗预测值与实验值平均绝对误差（MAE）和均方根误差（RMSE）分别为3.645W·h和10.421W·h，柑橘温度预测值与实验值的MAE和RMSE分别为0.034℃和0.042℃，相比Transformer模型，能耗预测的MAE和RMSE最高分别下降41.43%和39.27%，柑橘温度预测的MAE和RMSE最高分别下降46.03%和28.81%。本研究可为柑橘冷藏过程温度波动与能耗动态感知与优化调控等提供可靠方法支持与参考。

关键词:柑橘;冷藏效率;时序预测模型优化;PatchTST;TiDE

Abstract:

The temperature fluctuation during the low-temperature storage process of citrus is a key factor that triggers quality and safety risks for the fruit and increases refrigeration energy consumption. Simultaneously, quality and energy consumption are crucial evaluation indicators for assessing the efficiency of citrus cold storage. Achieving dynamic predictions for both aspects can provide reliable support for scientifically anticipating and precisely optimizing citrus cold storage efficiency. In light of this, a citrus cold storage efficiency time-series prediction model was proposed based on PatchTST. Firstly, a basic PatchTST model was constructed based on the self-attention mechanism and the channel independent (CI) prediction method. Secondly, by integrating the basic PatchTST model with the covariate feature extraction module from the TiDE model, feature extraction for all sequences in the multivariate time series dataset was achieved, effectively improving the model’s prediction accuracy. Finally, quantitative analysis of the correlation between cold storage refrigeration parameters, energy consumption, and citrus temperature was conducted by using the Pearson correlation analysis method. This analysis helped determine the input parameters for the TiDE-PatchTST model. The model was then trained and tested with 5000 sets of experimental data, and its accuracy and superiority were compared and validated against other models like basic PatchTST and Informer. The results showed that the predicted cold storage energy consumption values of the TiDE-PatchTST model had average absolute errors (MAE) and root mean square errors (RMSE) of 3.645W·h and 10.421W·h, respectively. The MAE and RMSE for citrus temperature predictions were 0.034℃ and 0.042℃, respectively. Compared with Transformer model, the MAE and RMSE in energy consumption predictions were decreased by up to 41.43% and 39.27%, and in citrus temperature predictions, they were decreased by up to 46.03% and 28.81%. The research result can provide strong support for the dynamic perception and optimization control of temperature fluctuations and energy consumption during the citrus cold storage process.

Key words:citrus;cold storage efficiency;time-series prediction model optimization;PatchTST;TiDE