Research On The Deep Learning Based Energy Consumption Analysis For Buildings

The building energy consumption is increasing in a fast speed with the development of urbanization.The building energy saving has been set up as an important national strategy.Artificial intelligence is the significant theory and techonology of smart buildings.The artificial intelligence is the important theory and basic technology of buildings.Utilizing the machine learning method to analyze the building energy consumption data plys a crucial rule in the building energy saving and optimization and building management.The deep learning is a new machine learning method,has good capacity in the multi-layer feature extraction and non-linear approximation.The prediction of building energy consumption using deep learning methods obtains more accurate and effective results than the traditional machine learning methods.To explore the hidden knowledge and rules in the building energy consumption data,this paper integrates the information of knowledge and data and studies the building energy consumption analysis based on deep learning with the centers of several problems including insufficiency of available data,promotion of prediction accuracy and the exploring building energy consumption pattern of multi-buildings.The main researches of this paper are as follows:Firstly,to overcome the insufficiency of building energy consumption data in the newly built and energy-saving renovation buildings,this paper proposed a data generation method for the building energy consumption using Generative Adversarial Nets(GAN),and realizes the data enhancement in the building energy consumption modeling.The method firstly makes use of GAN to find the hidden distribution of the original data and generate the parallel data via GAN.Then,after the data filtration,the mixed data set is formed via combining the original data and the artificial data.Finally,the mixed data is utilized to construct the prediction models.This method is applied in the commercial and office buildings,and the detailed experiments and comparisons are conducted.Experimental results show that the proposed method performs best compared with the existing data generation methods,and the mixed data driven prediction models perform better than the original data driven models.Secondly,for the purpose of promoting the prediction accuracy of building energy consumption models,this paper proposed a new hybrid method for the prediction of building energy consumption based on cyclic features and Deep Belief Network and Extreme Learning Machine based ensembled model,named(DEEM+CF).In the DEEM+CF,firstly the cyclic features are extracted by spectrum analysis,and the stochastic components are obtained.Then,the stochastic component is approximated by the new DEEM.Finally,the predicted stochastic components are combined with the cyclic features to obtain the final predicted results.Furthermore,the energy consumption prediction experiments are conducted in the commercial and office buildings.The experimental results demonstrated that the utilization of cyclic features and the feature extraction of each layer in the deep learning model can promote the building energy prediction accuracy effectively.Finally,to avoiding considering building physical parameters,weather,etc.in analysing the energy consumption profiles of multiple buildings and to mine the energy consumption profiles effectively,this paper proposed a data-driven method combining Density-Based Spatial Clustering of Applications with Noise and deep-learning to analyze the energy consumption profile of the multiple buildings.In this method,firstly,the DBSCAN is adopted to analyze the building energy consumption series.Then,some buildings in each building type are selected as the representative buildings,and the DEEM+CF is adopted to predict the daily energy consumption profile of the selected buildings.Finally,the predicted energy consumption profile in each type are averaged to be the predicted daily energy consumption profile of each type of buildings.The energy consumptions of residential building sector are selected to prove the availability of the proposed method.Experimental results prove that the dynamic building energy consumption profile can be effectively mined using the proposed method.