Nonintrusive Device-Level Pattern Recognition and Power Consumption Prediction in Buildings Through Sequential Variational Autoencoders

Abstract

Accurate forecasting of appliance-level power consumption is essential for advancing energy efficiency in smart grids and improving load management systems. However, simultaneous appliance operation, diverse usage patterns, and limited access to appliance-level data introduce significant challenges for effective load forecasting. This research presents a novel dynamic state-space model (SSM) that disaggregates household energy consumption at the device level, addressing these complexities. The model integrates temporal dependencies and inter-device interactions, capturing the stochastic and dynamic nature of device behavior over time. A sequential variational autoencoder (VAE) framework is employed to estimate hidden device states, overcoming limitations of conventional methods that struggle with complex transitions and long-term dependencies. We introduce enhancements to the VAE framework, including injective LSTM-based encoders, regularization techniques, and customized loss functions that ensure interpretability and improve identifiability of appliance states. This research has broad implications for sustainable energy applications, including energy efficiency optimization, real-time load management, and energy forecasting.