Research On Building Energy Consumption Prediction Based On Fast Attribute Reduction Of Weighted Neighborhood Rough Set

In recent years,with the continuous improvement of urban industrialization and electrification,the demand for energy has reached a new high.Taking electricity consumption as an example,it is expected that global electricity demand will increase to 1.5to 1.8 times the current level by 2050.The conservation and efficient use of building energy has become a focus of attention,and accurate and fast prediction of building energy consumption can guide the production and dispatch of energy,and is of great significance for the supervision of building energy consumption and energy conservation.The factors that affect building energy consumption can mainly be divided into three categories: time factors,meteorological factors,and building attributes.Among them,time factors are the most important energy consumption influencing factors,and building energy consumption is a typical time series.Time attributes include the time of electricity meter reading and holidays.Meteorological factors include temperature,cloud cover,wind speed,and wind direction.Building attributes are inherent factors that affect building energy consumption,including building area,floors,external wall materials,and use.In the context of big data,the prediction of building energy consumption generally faces problems such as too many feature attributes,difficulty in classifying energy use patterns,and generally low prediction accuracy.This paper proposes a fast attribute reduction algorithm based on weighted neighborhood rough sets and a new energy consumption prediction model based on a combination of rolling time domain and long-short term memory(LSTM)neural network.The main work is as follows:(1)Daily energy consumption data from 48 buildings and 316 equipment units from November 2020 to September 2021 were collected to establish an energy consumption feature data set.The collected data was preprocessed through deduplication,null value analysis,and outlier analysis to reduce the impact of nonlinearity and non-stationarity of energy consumption data on prediction performance.The energy use patterns were obtained through a combination of clustering and classification,forming a decision set.(2)A fast attribute reduction algorithm based on weighted neighborhood rough sets is proposed to achieve attribute reduction by judging the importance of each attribute in the energy consumption dataset.The performance of the algorithm is demonstrated by comparing it with the attribute reduction algorithms based on information entropy and neighborhood rough sets,and the experimental results show that the proposed algorithm based on weighted neighborhood rough sets has better performance,which can effectively avoid the influence of unreasonable input variables on prediction accuracy in big data scenarios.In the actual energy consumption prediction experiment of a building in Yunnan,the attribute reduction algorithm proposed in this paper screened out 15 energy consumption feature attributes,and achieved a classification accuracy of 94.34% with a reduction of 11.76%of training data,with the highest feature classification accuracy and energy consumption prediction accuracy.(3)A method for establishing an energy consumption prediction model based on LSTM with moving horizon is introduced,which introduces the feature of time-series data into the algorithm,and uses moving horizon technology to take the historical time-domain data as input features for practical building energy consumption prediction.The experimental results show that the LSTM prediction method with moving horizon reduces the root mean square error(RMSE)value by an average of 33.08% compared to the traditional non-moving method,with an average increase of 5.25% in training speed,resulting in better prediction accuracy and faster training speed,which significantly improves the accuracy and real-time performance of the energy consumption prediction model.(4)A front-end and back-end separated building energy consumption big data mining platform was designed based on Airflow,Flink framework,and database technology,which realizes the complete energy consumption prediction function of raw data input,pre-processing,and energy consumption prediction.The platform consists of four layers:presentation layer,application layer,service layer,and data layer,with functions such as analysis tools,data management,algorithm management,model management services,and data interface services.Experimental results show that this prediction platform can quickly and accurately predict building energy consumption,with strong robustness and stability,and can provide theoretical basis and method support for refined management of building energy consumption and decision-making management of building energy saving and emission reduction.