基于Transformer神经网络的建筑能耗预测模型

Building Energy Consumption Prediction Based on Transformer

建筑能耗预测是建筑能耗研究的重要方向之一,对于建筑节能以及建筑设备高效运行具有指导价值和必要作用。因此建筑能耗预测是节约能源与低碳环保要求大背景下的重要研究部分。设计了一种多输入多预测输出的基于Transformer神经网络的建筑能耗预测模型,利用Transformer网络对于时间序列数据分析能力对数十个不同的建筑能耗影响因素进行计算,并得到不同类型能耗情况。通过对气候条件、建筑物特性、设备使用行为等因素进行能耗相关性分析与数据标准化操作,利用整理后的数据驱动预测模型进行特征分析,最终获得各项能耗与各个因素间的预测函数。此外,通过对划分的测试集时间序列进行实验,得到其温度决定系数R^(2)、各项能耗决定系数R^(2)、温度MSE、各项能耗MSE分别为0.98080、0.88113、0.000211、0.00089。验证了模型的预测效果,证明了模型的先进性。

Forecasting building energy consumption is a crucial area of research in building energy studies,as it provides valuable guidance and significance for optimizing the operation of building energy systems and equipment.In the context of sustainability,especially with the growing emphasis on environmental concerns and carbon neutrality requirements,accurate building energy consumption prediction is an essential research component.To address this,a novel building energy consumption prediction model was developed using a Transformer neural network with multiple inputs and multiple predictive outputs.The Transformer network’s inherent ability to analyze time series data was leveraged to consider ten distinct factors that influence building energy consumption,allowing for the prediction of various energy consumption types.Through the analysis and standardization of energy-related factors such as climate conditions,building characteristics,and device usage patterns,the data was used to drive the prediction model for feature analysis.Ultimately,this process yielded the prediction function for each energy consumption item and its relationship with the influencing factors.To validate the model’s effectiveness,experiments were conducted on the time series data of the test set.The results demonstrated the superiority of the model,as indicated by the temperature determination coefficient(0.98080),the coefficient of each energy consumption item(0.88113),the temperature mean squared error(MSE)(0.000211),and the MSE of each energy consumption item(0.00089).The validation outcomes confirmed the model’s superiority in accurately predicting building energy consumption.