Research On Data Mining And Prediction Method Of College Building Itemized Energy Consumption Based On LSTM

At present,how to further implement energy conservation and emission reduction in the global response to climate change has attracted more and more attention.As one of the three giants of energy consumption,the building has become the main target of energy conservation and emission reduction strategy.Relying on the large-scale deployment of the energy conservation supervision system construction project of the Ministry of Housing and Urban-Rural Development of the PRC,college buildings gain a first-mover advantage in the collection of itemized energy consumption data of college buildings.Building itemized energy consumption refers to the itemized measurement or the different categories or ways of energy consumption on the basis of the total energy consumption.It will make the using ways of energy consumption more clear.However,most of the itemized energy consumption data collected by college buildings have not yet formed a standardized public dataset.The itemized energy consumption of demand information of these buildings is still in an unknown state.Therefore,a method for predicting hourly short term itemized energy consumption of a college building based on Long Short-term Memory(LSTM)is proposed in this thesis.The research is conducted on an administrative office building in a technological university.The main work of this thesis as follows:(1)It constructs a college building itemized energy consumption prediction model based on LSTM by the use of preprocessed building itemized energy data in this thesis.The model is validated and quantitative evaluated on the test set which is consisted of four seasonal periods called autumn,winter,spring and summer respectively.The result shows that the LSTM model basically solves the itemized energy consumption prediction problems on the autumn and summer seasons.But there are still some shortcomings on the winter and spring.(2)Based on the shortcoming of LSTM model and the inherent characteristics of time series of building energy consumption,it puts forward to using a clustering method called toeplitz inverse covariance-based clustering(TICC)to slice the energy consumption into different sequences.It will also compare two kinds clustering methods called K-Means and GMM to validate the efficient of TICC method.In the meantime,the results are evaluated by the DTW distance function to give a quantitative comparison among the three methods.The results show that the sequences sliced by TICC method are the most independent,which offers more refined data characteristics to the next predicting models.(3)Two itemized energy consumption models called TICC-LSTM and TICC-AR-LSTM are also constructed in this thesis.It also designs the framework and process of each prediction model,and makes the experimental verification and quantitative assessment on the four test sets described above.These two models are also compared with the LSTM model.The results show that the MSE and MAPE evaluation index of the LSTM model on the spring lighting energy consumption are 5.99 and 12.12% respectively,the summer air conditioning energy consumption are 1998.06 and 45.05% respectively,the all seasonal power consumption is 0.07 and 17.69%,0.01 and 27.46%,0.04 and 19.63%,0.04 and 16.01% respectively.The evaluation index of the TICC-LSTM model on the summer and autumn lighting energy consumption are 8.83,12.74%and 10.65,14.7% respectively.The evaluation index of the TICC-AR-LSTM model on the winter lighting energy consumption are 0.25 and 7.01%,the autumn,spring and winter air-conditioning energy consumption are 1.25 and 30.49%,51.09 and 28.88%,0.3 and 8.88% respectively.The models referred above achieves the best performance among the different sequence and itemized energy consumption.A general predicting method for college buildings itemized energy consumption is designed and a corresponding optimal model selection for different periods is provided in this thesis.The experimental results show that this method can effectively solve the problem of predicting of colleges building itemized energy consumption and provide reference value for the prediction task of college buildings which have not installed the itemized energy consumption acquisition system.