| Behind China’s rapid economic growth,energy consumption has become increasingly serious.Solving the problem of energy consumption is an indispensable issue in the process of social development.Among them,the total energy consumption of buildings accounts for a relatively high proportion of the total energy consumption in the China,and with the rapid development of urban construction,the energy consumption of buildings is still showing a trend of rapid growth.Energy consumption control is an effective way to alleviate building energy consumption,and it is also the most direct way to achieve energy saving.Effective energy consumption control is premised on prediction,so the improvement of energy consumption prediction performance has become a hot spot in the field of energy conservation.However,the realization of energy consumption prediction requires the collection of energy consumption data in different states,and the acquisition of large-scale energy consumption data is affected by many factors and high collection costs.Therefore,how to improve the predictive performance of building energy consumption is a major research task in the field of building energy efficiency.This article applies reinforcement learning methods to building energy consumption prediction,and effectively solves many problems faced by collecting energy consumption data through cross-building transfer.It is mainly divided into the following three parts:(1)Aiming at the problem that new buildings and newly installed instruments have smaller historical data sets and cannot accurately predict building energy consumption,a reinforcement learning building energy prediction method based on feature adjustment and cross-building transfer is proposed.This method uses a large amount of source building energy consumption data and a small amount of target building energy consumption data through the adjustment of seasonal and trend to process features and label values.It is used in the pre-processing and post-processing stages of reinforcement learning algorithms.Finally,use the adjusted data as input to the Q-Learning algorithm to achieve knowledge transfer across buildings.Through comparative experiments using data from four schools provided by Powersmiths of Canada,the results show that the proposed algorithm effectively improves the prediction accuracy of new buildings with limited data.(2)Aiming at the problem that the predictive performance of building energy consumption is affected by the uncertainty of many complex factors,and the labeled data set is not always available,an energy consumption prediction method and algorithm for reinforcement learning based on feature extraction and cross-building transfer is proposed.The main research is to provide energy prediction solutions for buildings with limited historical data by using the energy consumption data of other types of buildings when the data is not labeled.Firstly,the Conditional Restricted Boltzmann Machine is used to automatically extract advanced features for state assessment,and the building energy consumption is modeled with the reinforcement learning algorithm.Secondly,the obtained continuous state is used as the input of the reinforcement learning algorithm.Finally,by comparing the energy consumption data recorded by Baltimore Gas and Electric Power Company of the United States,experiments show that the proposed method effectively improves the predictive performance of energy consumption.(3)Aiming at the problem of insufficient training samples in the prediction of building energy consumption,which leads to a slow learning speed in the target field,an energy consumption prediction method based on reinforcement learning based on generative adversarial networks and cross-building transfer is proposed.This method employs unsupervised manifold alignment to directly adjust and reuse knowledge mapped from the source task to the target task,and combines the adaptive control algorithm to obtain the optimal policy in the target domain.Furthermore,the algorithm introduces generative adversarial networks to reduce its dependence on real samples in the target domain.Finally,it is used to transfer between the classic Mountain Car and Inverted Pendulum problems in reinforcement learning,and compared with the energy consumption data of different buildings in Chapter 4.The experimental results show that the proposed algorithm can effectively improve the use of samples and speed up the learning speed in the target domain,have faster convergence speed and higher stability. |
