EDA-LSTM Model Based On Research On Building Energy Consumption Prediction Method

It is particularly important to save energy in the context of global resource shortage and environmental deterioration.China is the world’s largest energy consumer which its building energy consumption ranks first.At present,China adopts green energy-saving technologies to try to reduce the energy consumption of buildings.However,with the continuous progress of energy-saving technologies,the potential for energy-saving transformation at the technical level is decreasing.In comparison,building energy consumption prediction is an alternative way of improving building energy efficiency.Building energy consumption prediction can understand the energy consumption of buildings,predict operating costs,and determine appropriate energy-saving measures,so as to provide decision support for the diagnosis and renovation of building energy efficiency,so as to reduce building energy consumption.Therefore,this paper selects a green office building in Shenzhen as the research object,and starts with optimizing and improving the prediction model methods such as data preprocessing,establishing a combination model and improving a single model,so as to improve the prediction accuracy of the model and provide more accurate decision support for the diagnosis and transformation of building energy conservation.Firstly,this article summarizes the common methods of building energy consumption prediction,the adaptive scenarios of various methods,and their advantages and disadvantages.The principles of the Decision Tree,Support Vector Machine,Neural Network and other models used in this article are described,and the algorithms needed to optimize the algorithm model based on the characteristics of building energy prediction are introduced.Secondly,according to the idea of establishing the combination model,a combination model CART-SVR which combines the CART algorithm and theSVR algorithm is constructed,and a combination framework is designed to explain the model structure of the combination model.The data was derived from the building energy consumption of a green building in Shenzhen with an interval of one hour in the year of 2018.The data was preprocessed by data cleaning,data integration,data normalization and other preprocessing technologies,and the feature selection was made by using correlation analysis.Finally,the training network of the algorithm is established,the parameters of the combined model were determined,and the experiment was compared with the single models SVR and LSTM is performed to verify the effectiveness of the combination model.The prediction accuracy of the combination model was 97.84%,higher than that of the SVR and LSTM models.In addition,the prediction error of CART-SVR is lower than that of the comparison model.Therefore,the combination model CART-SVR has higher prediction accuracy,which demonstrates the advantages and feasibility of the combination model.And it can be used as an effective method for the prediction of building energy consumption.At the same time,eight operating modes of green building was identified.The study found that the importance of factors influencing building energy consumption was listed as follows: air conditioning terminal energy consumption,water supply and drainage pump,garage power,energy consumption of heating and cooling stations of central air conditioning,other energy consumption,energy consumption of elevator,energy consumption of lighting,energy consumption of tenants,energy consumption of blower,and CO2.Finally,the model EDA-LSTM for Long short Time Memory based on double attention mechanism is constructed.The parameters of the attention mechanism layer are selected by the competitive random search algorithm.The data preprocessing adopts the feature combination method to introduce the duality.The EDA-LSTM model was compared with LSTM,MLP,Light GBM and SVR to verify whether the EDA-LSTM model was better than other common models,and the prediction error was evaluated.The MAE of the EDA-LSTM model is 4.02,which reducing 22.43-47.89.The RMSE of the EDA-LSTM modelis 2.87,which decreases by 8.99-28.14.The experimental results show that the EDA-LSTM model is more suitable for the prediction of building energy consumption.In addition,comparing the SVR model with the duality and the conventional pre-processed SVR model,it is found that the duality can significantly improve the prediction accuracy of the model.At the same time,the study found that in addition to the single feature,the binary feature is also an important factor affecting building energy consumption.For example,the binary characteristics of air-conditioning terminal and working day type,hour type,CO2,garage power,lighting,elevator,blower,tenant and other energy consumption combinations;The binary characteristics of energy consumption and CO2 of cold and hot station,working day type,hour type,water supply and drainage pump,lighting,elevator and garage power combination,etc.Both the prediction models CART-SVR and EDA-LSTM proposed in this article can improve the prediction accuracy of the model.And the prediction models CART-SVR and EDA-LSTM provide new options for researchers to predict building energy consumption.In the future,deep reinforcement learning technology can be applied to the prediction of building energy consumption to obtain higher prediction performance.