基于高阶特征挖掘和贝叶斯MCMC方法的建筑能耗预测
Prediction of building energy consumption based on advanced feature mining and Bayesian MCMC method
摘要：
针对建筑能耗预测模型需要解决解释变量选取困难、建筑样本数据有限和建模算法难以捕捉复杂的非线性关系等问题,提出了基于高阶特征挖掘和贝叶斯MCMC(马尔可夫链蒙特卡罗)方法的能耗预测模型。本研究利用建筑本体、建筑环境及建筑人员活动等特征指标作为模型输入,以CBECS数据库中205栋办公建筑的供热能耗数据作为训练集样本,选用符号检验与K-fold交叉验证识别高响应交互作用因子组合,通过MCMC方法进行重要性采样,预测了办公建筑的供热能耗强度。研究发现,建筑的层高、面积、层数之间存在高阶交互关系,模型预测精确度指标归一化平均偏差(NMBE)为3.5%,均方根误差(RMSE)为0.014,均方根误差变异系数(CVRMSE)为14.0%,可有效预测办公建筑供热能耗。
Abstract：
To solve the problems of difficult selection of explanatory variables, limited building sample data and difficult capturing of complex nonlinear relationships in modeling algorithms, an energy consumption prediction model based on advanced feature mining and Bayesian MCMC method is proposed. In this study, the characteristic indicators such as building ontology, building environment and occupancy behavior are used as model inputs to identify high-response interaction factor combinations, and the heating energy consumption data from 205 office buildings in the CBECS database are used as training set samples, and the combination of high-response interaction factors is identified by symbol test and K-fold cross-validation, and the importance sampling is carried out by the MCMC method to predict the heating energy consumption intensity of office buildings. It is found that there is a high-order interaction between the floor height, area and number of floors of the building, and the normalized mean biased error (NMBE) of the model prediction accuracy index is 3.5%, the root mean square error (RMSE) is 0.014, and the coefficient of variation of the root mean squared error (CVRMSE) is 14.0%, which can effectively predict heating energy consumption of office buildings.
Keywords:advanced feature mining; K-fold cross-validation; Bayesian model; MCMC method; energy consumption prediction
