A Data-Driven Digital Twin Framework Using Long Short-Term Memory Networks For Intelligent Energy Optimization In Residential Buildings

Abstract

With the increasing emphasis on sustainability and energy-conscious living, there is a growing demand for adaptive systems that can intelligently manage residential energy usage. This study proposes a data driven approach that combines Long Short-Term Memory (LSTM) neural networks with Digital Twin (DT) technology to support long-term energy forecasting and optimization in non-automated residential buildings. The LSTM model was trained using real-time data collected from multiple homes, focusing on variables such as indoor temperature, occupancy, humidity, and electricity consumption. Corresponding DT models were built to simulate the physical behavior of each building and validate the predictive performance under various operational scenarios. The hybrid system achieved high accuracy in forecasting, with a Mean Absolute Percentage Error (MAPE) under 7% and R² values consistently above 0.91. Additionally, simulation-based energy control using this model demonstrated annual savings between 19% and 21%. The study builds on prior ANN-based research and introduces an adaptable and scalable methodology suitable for conventional residential contexts.