Innovative Deep Learning Techniques for Energy Data Imputation Using SAITS and USGAN: A Case Study in University Buildings

Abstract

The growing demand for efficient energy management in buildings highlights the critical need for accurate and robust data. This study explores innovative approaches for multivariate time series imputation in the context of electrical energy consumption using advanced deep learning models: Self-Attention Imputation for Time Series (SAITS) and Unsupervised Generative Adversarial Network (USGAN). The novelty of this work lies in the evaluation of these models for imputing active power, apparent power, and power factor within the electrical loads of the Universidad de Las Américas (UDLA) in Ecuador. Additionally, the models were tested on different levels of random data loss: 10%, 20%, 30%, 40%, and 50%. Results from this study demonstrate the effectiveness of these models in imputing missing values across various percentages of data loss. These findings contribute valuable knowledge for optimizing energy management and promoting sustainability in educational institutions, serving as a benchmark for similar environments.