面向密闭空间内外温度的时序预测模型

Time Series Prediction Model for Internal and External Temperature of Confined Space

目的 研究密闭空间条件下已知外部温度时间序列对内部实时温度的预测推理问题。方法 选取密闭空间内外温度时序预测典型场景,抽象为多变量时间序列预测问题,分析变量间的关联性和依赖性。借鉴特征融合、注意力机制、多任务模型等思路,结合物理机制与数据特征,基于长短期记忆网络基本网络单元,构建密闭空间内外温度时序预测模型,并在万宁、敦煌、漠河对某型密闭空间进行数据采样,基于三地数据集进行不同模型试验。结果 多变量模型比单变量模型具有更好性能,注意力机制对该场景没有显著性能提升,结合物理机制的模型结构设计充分考虑了变量之间的关联性和依赖性,能显著提升预测精度,双输入双输出的多变量时序预测模型具有相对最高的精度和最稳定的鲁棒性,是面向密闭空间内外温度时序预测的相对最优模型。结论 研究结论可指导密闭空间其他环境特征建模,研究思路可为其他多变量时序建模问题中变量之间的关联性、依赖性分析提供参考。

ed as a multi-variable time series prediction problem,and the correlation and dependence analysis among variables were the key difficulties.By referring to the ideas of feature fusion,attention mechanism and multi-task model,combined with the physical mechanism and data characteristics,and based on the basic network unit of long and short term memory network,the internal and external temperature time series prediction model of confined space was constructed.The data of a certain type of confined space was collected in Wanning,Dunhuang and Mohe,and different model experiments were carried out based on the data sets of the three places.The multi-variable model had better performance than the univariable model,and the attention mechanism did not significantly improve the performance of this scenario.The model structure design combined with the physical mechanism fully considered the correlation and dependence between variables,which could significantly improve the prediction accuracy. The multi-variable time series prediction model with double input and double output had the highest accuracy and the most stable robustness. It was a relatively optimal model for the prediction of internal and external temperature time series in confined space. The research conclusions can guide the modeling of other en-vironmental characteristics in confined space, and the research ideas can provide references for the correlation and dependency analysis among variables in other multi-variable sequential modeling problems.