Data-driven Energy Consumption Prediction Of Building HVAC System And Model Optimization

Heating,ventilation,and air conditioning(HVAC)system is one of the major energyconsuming devices in buildings,and accurate prediction of HVAC system's energy consumption is of great practical significance for building energy conservation and management.Data-driven energy consumption prediction methods have been widely used due to their advantages of model building simply and high accuracy.This paper carries out the research of energy consumption prediction for an office building using support vector regression algorithm.The water source heat pump system is used for winter heating and summer cooling for the building.The local meteorological parameters,the HVAC system operational parameters and the indoor air temperature are employed to establish the models.In addition,the models are optimized by the methods of outlier detection,feature selection and parameter-tuning.On the basis of the local meteorological data and the HVAC system operational data under the summer cooling conditions,and the boxplot,the local outlier factor and the PCOut methods are separately used to detect the outliers in the original data set.Through analysis of the testing results of the energy consumption prediction models and the distributions of abnormal samples,the outliers in training data set have great negative impact on the prediction performance of data-driven models.And the possible outliers can be found by the methods of boxplot and local outlier factor,whereas the PCOut method fails to detect the potential outliers.Further research has found that the abnormal samples that detected by the boxplot method and the local outlier factor method simultaneously may be the most possible outliers in the training data set.Compared with the model that trained by the original data set,the root mean square error of the model that optimized by the outlier detection methods falls to 2.76 from 6.44 and the relative square error falls to 0.11 from 0.58 during testing course.From the above,the model performance for predicting energy consumption has been improved significantly by the outlier detection methods(i.e.,boxplot and local outlier factor).On the basis of the local meteorological data and the indoor air temperature under the winter heating conditions,and two variables,day type and time type,are introduced according to the energy consumption characteristics of office buildings.Then 6 energy consumption prediction models are established by the combinations of different input variables.According to the error analysis of model testing results and the sensitivity analysis of input variables,the parameters of illumination intensity,day type and time type have a positive effect on the model performance,and the sum of their relative importance is more than 70%.Hence,these three variables play a leading role in the model and are the optimal input feature.To further optimize the model,10-fold cross validation and grid-search methods are used to find the optimal values of SVR-based model parameters C and ?.Results show that the root mean square error of the model falls to 4.1160 from 7.8227 and the relative square error falls to 0.1194 from 0.4313 during testing course.