A deep learning approach to energy disaggregation using TCN layers for appliance and load insights

Abstract

Non-Intrusive Load Monitoring (NILM) plays a vital role in energy efficiency by disaggregating appliance-level consumption from aggregated household energy data. This study explores the use of Temporal Convolutional Networks (TCNs) for parallel appliance classification and load prediction, addressing challenges like overlapping energy signatures and long-term temporal dependencies. TCNs, with their dilated causal convolutions and efficient parallel processing, are well-suited for NILM applications, offering improved scalability and accuracy over traditional machine learning and recurrent neural network (RNN) approaches. The proposed framework utilizes multi-task learning to classify active appliances and predict their energy consumption simultaneously, reducing computational overhead and enhancing system adaptability. Experiments on publicly available datasets REDD, UK-DALE, demonstrate the TCN model's superior performance, achieving higher classification accuracy, improved load prediction fidelity, and robustness under noisy conditions. The lightweight and scalable architecture ensures suitability for real-world deployment, including smart grid systems and residential monitoring.