Forecasting building energy demand and on-site power generation for residential buildings using long and short-term memory method with transfer learning.

This study evaluates the effectiveness of the long and short-term (LSTM) implementation with a particular emphasis on assessing the impact of transfer learning techniques in improving prediction accuracy for building energy demand and on-site power outputs using empirical data from real-world building environments. The initial study utilized simulated data from a single-family prototype building model, employing cluster analysis to segment the training and testing datasets based on distinct cooling and heating periods. Subsequently, real-world data from an existing residential building was incorporated by utilizing LSTM-based transfer learning to improve the prediction accuracy of building energy demand and on-site power generation within a target domain. The training and testing phases involved pre-processed datasets with distinct time-series datasets for environmental, electricity demand, and on-site power generation data. The input variables in the architecture of the machine learning model included environmental, time-related data, and past-day energy datasets. This study also implemented interquartile range (IQR) analysis during the data pre-processing phase to effectively bridge the gap between the source and target domain feature and label spaces to minimize discrepancies and improve the accuracy of prediction performance. The results showed the LSTM model initially developed for a source domain effectively predicted energy demand and on-site power generation across summer and winter. Within target tasks, while initial transfer learning enhancements in prediction accuracy were modest due to each domain's low relevancies in their features and labels, significant improvements were achieved following strategic data pre-processing. The results underscored the importance of detailed pre-processing analysis in LSTM-based transfer learning models for accurate energy demand forecasting in real-world settings. The integration of transfer learning with IQR analysis refined the prediction capabilities of the models under practical conditions. • LSTM-based model enhances energy prediction accuracy. • Transfer learning enhances accuracy in on-site power predictions. • Cluster analysis segments training/testing in simulated data. • Real-world data integration improves LSTM predictions. • IQR analysis bridges gaps, enhancing prediction accuracy. [ABSTRACT FROM AUTHOR]