Research On The Deep Learning-based Method For Building Energy Consumption Prediction

Energy consumption prediction is essential for building optimization design,optimization control,demand-side response,and energy audit.It is of great significance for improving the energy efficiency of buildings and achieving the “dual carbon” goal in the building sector.Datadriven methods are popular in the field of energy consumption prediction.They are easy to use and have high accuracy.With the rapid development of information technology and artificial intelligence technology,massive building operational data and various algorithms make datadriven methods have greater development potential.Building operational data has temporal and spatial characteristics.In the temporal dimension,building operational data as time series have periodicity and time lag.In the spatial dimension,similar buildings often have similar energy consumption patterns.Their operation time series may be similar.At present,domestic and foreign researches still have shortcomings in these two dimensions.In the temporal dimension,most prediction models disorganize the time series.How to mine the characteristics of time series to further improve the prediction accuracy is a difficult problem that has not been solved currently.In the spatial dimension,there is still a lack of effective solutions for flexibly using the existing data of multiple similar buildings to help the development of prediction models in a few-shot scenario.To solve the above shortcomings,this paper focuses on the temporal and spatial characteristics and uses deep learning algorithms to facilitate accurate and reliable building energy consumption prediction models.The main researches are as follows:(1)Aiming at the feature extraction problem in the temporal dimension,a time series prediction method based on recurrent neural network and backpropagation feedforward network is proposed,which aims to mine the temporal characteristics of building operational data and further improve the accuracy of energy consumption prediction.This method combines a recurrent neural network and a backpropagation feedforward network.The recurrent neural network is used to extract important features from time series,and the feedforward network is used to learn the complex relationship between features and energy consumption.This method couples feature extraction and energy consumption prediction in a model and achieves iterative optimization of two neural networks to minimize prediction errors.The method is applied to an actual public building,and the results show that the method can effectively extract features of time series and has higher prediction accuracy than traditional data-driven methods.(2)Aiming at the problem of energy consumption prediction in a few-shot scenario,a federated learning-based method for constructing building transfer networks is proposed,which aims to realize the flexible use of multi-party data in the spatial dimension.This method combines similar buildings to establish a transfer network,in which each building establishes a local prediction model,and aggregates them to form a federated model that simultaneously reflects the common energy consumption pattern of multiple buildings and the individual energy consumption characteristics of a single building.Data sharing and model transferring are completed in the building transfer network through global training and local fine-tuning of the federated model.The method is applied to the actual buildings,and the results show that the method can make full use of multi-party data to improve prediction accuracy,decrease model training time by more than90%,and solve the energy consumption prediction problem in few-shot scenarios(3)To improve the application reliability of data-driven models,a general interpretation process for energy consumption prediction models is proposed,which aims to interpret the prediction results and provide decision support for building energy management.The process uses the sensitivity index to measure the impact of model inputs on the energy consumption prediction value,uses the weighted Manhattan distance to measure the difference of operational conditions,and mines the model prediction error change trend based on the relationship between the error and the distance.The process is used to interpret various data-driven models,and the results show that the method can interpret prediction models from multiple aspects,and evaluate the reliability of models.The method system established in this paper deeply explores the temporal and spatial characteristics of building operational data,improves the prediction accuracy of data-driven models,realizes energy consumption prediction in few-shot scenarios,quantitatively evaluates model reliability.It helps data-driven models play a greater role in the refined management of building energy consumption.