Research On Energy Consumption Prediction Of Office Building Based On Deep Learning

In the development of building energy conservation and intelligent building,it has become one of important development orientation to adjust building operation strategies,to conduct energy management and building energy diagnosis based on the results of energy consumption prediction.The method of building energy consumption prediction based on deep learning is widely used for its high accuracy,high stability and simple modeling.In this thesis,an office building in Wuhan city is taken as the research object,and a deep learning based combinational model which combines the denoising autoencoder and long short-term memory network is established based on the simulation data to make a shortterm prediction of the energy consumption of the office building,finally obtain a prediction performance which has higher prediction accuracy and stability compared with other models.The original feature variables include environmental data,building operation data and other types of data,the empirical method is used to preliminarily eliminate the variables whose variable value is constant zero,the unguaranteed 50-hour design specification is used to identify the outliers of meteorological variables and replace them with the design specification values corresponding to the variables,and the boxplot is used to identify the running variables and other types of variable’s outliers and replace the outliers with the mean of the former and the latter to improve the quality of the feature data.Denoising autoencoder is used to conduct data dimensionality reduction processing for highdimensional feature variables,and 2 is taken as dimensionality reduction descending gradient,obtain a total of 18 feature representations group with different dimensions.The results show that when the feature dimension is reduced to 28 dimension,the performance of the reduced dimension model converges to 0.0007 on the prediction data set.Which means that the complexity of the prediction model is greatly reduced while the original feature data information is basically retained,and the convergence of model training and counter overfitting are effectively accelerated.Aim at the time-sequence characteristics of energy consumption feature data,the deep learning based prediction model long short-term memory network is built for a series of feature representations obtained by dimensionality reduction,and the energy consumption after one hour was predicted after model training and optimization.Three comparison论文models,such as independent model multivariate linear regression and long short-term memory network,combined model principal component analysis--long short-term memory network,were established to predict energy consumption based on original data.The prediction performance result of target combined model shows the change trend that with the increase of dimension reduction dimensions increase first and then decrease,and the target model has the best prediction performance when the dimension of dimension reduction is 22.Compared with other models,the root mean square error of the prediction result of the target model decreases by 7.26%～44.44%,the mean absolute percentage error decreases by 25.14%～82.23%,decision coefficient increases by 0.62%～7.79%,the deep learning based combinational model’s prediction accuracy and prediction stability are improved obviously,which means that the target combined model is able to be applied in office building energy consumption prediction work effectively.