| Energy consumption by buildings accounts for a considerable proportion of the total energy consumption.And within the building energy consumption,public buildings contribute the most.Thus,accurate prediction of building energy consumption is important for energy saving and improving economic efficiency.Echo state network(ESN)has become a research hotspot in the field of time series prediction due to their good dynamics and modeling efficiency.However,the traditional ESN is a single objective learning model with a single reserve pool layer consisting of an input layer,a hidden layer(reservoir)and an output layer.It has limited computational power and low robustness.So it is unsuitable for multi-building energy consumption prediction problems,limiting its wide application and development.In order to make use of the advantages of ESN and compensate for its shortcomings,this thesis adopts an improved model,namely Chain-structure Echo State Network(CESN)to predict the energy consumption of public buildings.To achieve this goal,the CESN algorithm is designed to meet the requirements of different energy consumption prediction tasks.The main research work and innovations of the paper are as follows.(1)When predicting energy consumption for multiple buildings,if independent modeling is adopted,the prediction speed decreases with the increasing number of buildings.To solve this problem,before proposing an energy consumption prediction algorithm for buildings,this thesis performs a clustering analysis of energy consumption data,explores the characteristics of energy consumption data in different categories of multiple buildings under different division methods,and analyses the classification results using supervised learning methods.By performing cluster analysis on building energy consumption data,the energy consumption data of multiple buildings can be replaced by a small number of cluster centers while retaining as many energy consumption variation features as possible.This process can effectively reduce the computational burden.(2)The CESN model that used for predicting energy consumption of public buildings is constructed by stacking the echo state networks in a logical way.When predicting a single building,CESN is expected to improve the accuracy and robustness of prediction by extracting the historical energy consumption characteristics of the building to be predicted,as well as the historical energy consumption characteristics of other related buildings.To verify this assumption,this thesis designs single-step energy consumption prediction experiments for a single building in both data-complete and data-absent conditions.In addition,weather information is incorporated into the single-step prediction experiments;and multi-step prediction experiment is also introduced.The results show that the CESN model proposed in this paper outperforms the classical ESN model and other classical machine learning models in terms of accuracy and robustness.(3)Combining the above two projects,this thesis further designs a multi-building energy consumption prediction algorithm based on CESN.In order for CESN to have both fast prediction speed and high prediction accuracy,it is necessary to set the appropriate number of ESN submodules as well as the number of clustering categories.For this purpose,a target building to be predicted is selected in this study,and the prediction errors are recorded under different numbers of ESN submodules.The appropriate number of ESN submodules was determined by observing the change curve of the number of ESN submodules and the prediction error.Similarly,in order to determine the appropriate number of clustering categories and clustering method,this thesis uses two clustering methods,i.e.K-Means and GMM,to form three ways of clustering,e.g."time-domain K-Means","time-domain GMM" and "frequency-domain GMM".By observing the changing curve of the number of clusters and the prediction error,the best clustering method and the appropriate number of clusters are determined. |
