Building Electricity Demand Forecasting Model Based On LSTM And Integrated Learning Strategy

With the goal of decarbonizing building electricity,the application of wind,solar and other types of renewable energy in the building industry is becoming increasingly popular.Due to the stochastic nature of new energy generation power,which can easily lead to grid instability and other problems,the accurate prediction of building power demand plays an increasingly important role in the balance of energy supply and demand,and peak demand response.Although many data-driven models have been widely used in energy consumption prediction,there is still a lack of short-term prediction models with high prediction accuracy and strong generalization ability.To address this problem,this paper uses deep learning methods to focus on data preprocessing,integrated prediction framework construction,and integrated pruning.The main research contributions of this paper include:(1)Analyzing building energy usage patterns and combining FCM+KNN data clustering techniques to classify data related to building energy consumption.The LSTM(Long Short Term Memory Network)model combined with parameter optimization algorithm is used to build a short-term building energy demand model,and its prediction performance is verified through real cases.(2)Eight sub-predictors are further selected to construct a parallel integrated prediction framework.The sublearners specifically include SSA-LSTM,TLBO-BP,SVM,PSO-ELM,RF,ANFIS,BLR,and WNN.based on which the optimal combination of sublearners is obtained by supervised integration pruning.In the data integration part,MLR(multiple linear regression)is used to assign weights to the sublearners.(3)The accuracy and generalization ability of the proposed integrated model are validated by three energy use cases at different scales.Among them,the validation cases are from the public data of the energy prediction competition with a time step of 1 hour,with many data features and good data integrity,which are used for the validation of the model accuracy;Case A is the electric energy data of Yizheng city,with fewer data features and a large time span,representing a long-term electric energy use scenario with a large energy scale;Case B is the actual measurement data of the computer science and engineering building of Jiangsu University,with a time step of 1 day,representing Case B is the actual energy consumption of a single educational building with a time step of 1 day.The results show that the integration strategy proposed in this paper achieves the optimal accuracy in all three cases.Among them,the integrated pruning strategy can further reduce the prediction accuracy(MAPE)from 1.92 to 1.90.