Transferability of ML Time Series Prediction for Energy Forecasting in Smart Homes

electricity consumption prediction, IoT, TCN, TCN-LSTM, smart grid

This work explores the use of machine learning models (ML) in the context of Internet of Things-enabled smart energy management systems, particularly focusing on home energy management systems (HEMS). With the growing adoption of such devices, these systems have the potential to improve energy efficiency and reduce costs. This paper examines the feasibility of using time series prediction models for energy consumption forecasting, replacing traditional methods like Auto-Regressive Moving Average (ARMA) with deep learning approaches, namely Time Convolutional Network (TCN) and Temporal Convolutional Network - Long Short-Term Memory (TCN-LSTM) architectures. Using two smart home datasets, NIST and IHEPC, the paper evaluates the transferability and accuracy of the models. Results indicate that while the models perform well within a single dataset, they struggle to transfer reliably between datasets, likely due to the limited feature set used. Despite this, the models can be deployed on low-power devices with artificial intelligence (AI) chips, though their real-world application may require significant investment in sensors or reliance on third-party Application Programming Interfaces. The findings highlight the potential of machine learning in smart energy systems, while also addressing challenges related to model transferability and practical deployment. These findings contribute to Smart Cities Modeling by highlighting the role of machine learning in optimizing energy use for sustainable urban systems.