Studies On Energy Consumption Benchmark Methods Of Public Buildings Based On Data Mining Techniques

Public buildings are the main contributor for building energy consumption in China.By adequately measuring building operating energy benchmark,cities can better evaluate building energy savings and relieve the increase in energy consumption of public buildings.There are two categories of building operating energy benchmark according to the main application scenarios: evaluating energy consumption performance of comparable buildings and previous energy consumption performance of the same building.The former aims to identify the buildings with higher energy saving potential,and the latter one aims to identify anomalous energy usage which are mainly due to incorrect operation behaviors and realize refined energy management based on the former.In this context,this study proposed two novel data-mining approaches corresponding to the two application senarios based on the operating energy consumption data of public buildings in Chongqing,which can provide deeper insights for building energy management during operation phase.The main contributions of this study are summarized as follows.There is still lack of a comprehensive and public-availiable operating energy consumption database of public building in China,and the current building energy consumption database is limted in the sample size and dimension of the information related to energy usage.Therefore,this study conducted an online investiagetion of operating energy consumption for large-scale public buildings in Chongqing.The data of annual energy consumption and the information related to energy usage for approximately 3000 public buildings were collected,which were the dataset prepared for establishing the benchmark approach for evaluating energy consumption performance of comparable buildings.Then,the statistical charateristics of energy consumption and the information data were analyzed based on descriptive statistical analysis.The results showed that the annual energy consumption and the most information data were heterogeneous for the buildings with the same function even though the data were transformed to natural logarithm.Besides,the annual energy consumption per squared meters were statisfied the requirements of the current energy standards for more than 80%of the total buildings.Measuring the effects of influencing factors of building energy consumption is an important step for establishing the energy benchmark approach for comparable buildings,which enables to explain the main reasons of the significant differences acorss the buildings.A regularization-based machine learining method named Group Lasso was introduced in this study to identify the main drivers in statistical models of building energy consumption.This approach was demonstrated to the collected dataset for hospital and education buildings considering almost 30 independent vairiables related to energy usage,including building service level,physical and equipment features and energyefficiency operating management.Meanwhile,the stepwise regression method was slectected to compare with Group Lasso.The results showed that Group Lasso reduced30%-40% of the number of independent variables compared with the stepwise regression while ensuring the model accuracy was basically unchaneged.This indicated that the key drivers identified by Gorup Lasso can be more representative.The contribution of the results is to balance the cost and complexity of data collection for constructing building energy consumption database in Chongqing.A benchmark approach for hospital and education buildings using quantile regression method was proposed based on the determined key drivers.A set of quantile regression models were built for eliminating the effects of the key drivers so that the energy consumption across the buildings can be comparable.Then a theoretical distribution of energy usage for each individual building was created.Finally,the approach proucded the distribution of scores for each building based on efficiency and determined the final score by comparing with the actucal energy usage.The proposed approach revealed the significance and effects of the key drivers were always varied at each quantile instead of remaining unchanged.Furthermore,the key drivers of energy consumption for the buildings in upper quantiles and lower quantiles were determined.The proposed approach can help policymakers identify the energy saving potentials for buildings and formulate energy incentive policies that encourage competition between buildings.To realize refined energy management for individual buildings,a data-mining based approach to identifying anomalous daily electricity load profiles was proposed.The approach first aims to identify the typical daily electricity load patterns(TELPs)by a twostep clustering method.Five statistical features were defined to reduce the dimension of raw daily electricity load data.Then the DBSCAN algorithm was applied to extract and remove the outliers of daily electricity load profile.The k-means algorithm was used to identify the TEPLs.On this basis,the classification tree of TELPs and potential influencing factors was built based on CART decision tree,which reveals that how the factors affect daily electricity load pattern.The proposed approach was validated on the historical electricity load data for three practical office buildings.The results showed that the proposed approach enabled to effectively identify TELPs and predict the expected pattern for any given day.The prediction accuracy was more than 80% for all case buildings.This approach can provide building managers with an efficient way to understand the characteristics of building electricity usage patterns and detect anomalies therein.