面向有限传感器的建筑冷负荷深度学习预测模型构建

Construction of deep learning prediction model for building cooling load with limited sensors

建筑冷负荷准确预测是实现建筑空调系统基于模型的优化控制和建筑节能的重要基础。面向实际建筑建立准确的冷负荷预测模型是一个十分困难的系统工程,包括传感器安装、数据的通信和读取、数据预处理和模型识别。在实际工程中,传感器安装数量和种类有限或者运行过程被损坏使得获取必要的模型识别数据十分困难,同时实际建筑复杂的现场情况使得建立物理模型困难。针对以上2个问题,本文构建了面向有限传感器的建筑冷负荷预测深度学习模型,并结合实际案例,分析了冷负荷的主要影响因素,采用传感器替代、补充和数据插值的方法构建模型输入集,利用长短期记忆模型建立了高性能的冷负荷预测模型,并分析了不同特征量输入集下的模型性能。在烟厂厂房的冷负荷预测中,利用本文构建的冷负荷预测模型识别流程建立的长短期记忆模型的平均绝对百分比误差(MAPE)为9.09%。

Accurate prediction of building cooling load is a crucial foundation for implementing model-based optimization control of HVAC systems and achieving energy efficiency in buildings.Establishing an accurate cooling load prediction model for actual buildings is a challenging engineering task,including sensor installation,data communication and reading,data preprocessing and model identification.In practice,it is difficult to obtain the necessary model identification data due to the limited number and variety of sensors installed or being damaged during operation,and the complex site conditions of the actual building make it difficult to establish a physical model.In order to solve the above two problems,this paper develops a deep learning model for predicting building cooling load with limited sensors.Through an actual case study,this paper analyses the main factors affecting cooling load,constructs the model input set by using the methods of sensor substitution,supplementation and data interpolation,establishes a high-performance cooling load prediction model by using the long short-term memory model,and analyses the model performance under different feature inputs.In the cooling load prediction for a tobacco factory workshop,the mean absolute percentage error(MAPE)of the long short-term memory model established by the recognition process of the cooling load prediction model constructed in this paper is 9.09%.