A Hybrid Model for Multi-Step Prediction of Building Energy Load Based on EWT Extraction and LSTM

Abstract

Accurate building energy load multi-step prediction facilitates building energy efficiency and management. Time series analysis is often used to deal with the forecasting problem, and the commonly used forecasting algorithm is the Long Short Term Memory (LSTM) algorithm. As building energy consumption data is generally seasonal, uncertain and non-linear, it is difficult to achieve high accuracy with a single algorithmic model. In this paper, the empirical wavelet transform (EWT) method is used to extract the eigenmode functions (IMFs) from the data to overcome the seasonality and non-linearity of the data. However, the EWT approach suffers from the problem of edge effects and the obtained IMFs components have boundary errors at the end, which arise due to the definition of EWT. In this paper, a hybrid model is designed to correct for the boundary error in the EWT decomposition, and the results show that the modal components obtained from the improved empirical wavelet transform are more suitable as input data for the prediction model. Based on the building operation data obtained from real measurements, accurate prediction results can be obtained and can be used as a reference for related prediction studies.