Online Load Monitoring in IoT Residential Buildings: Deep ResNet and CamAL Fine-Tuning for Accurate Appliance State Prediction

The need to monitor individual appliance states in a smart home is driven increasingly by the demand for energy efficiency and automated home management. However, widespread adoption is often hindered by the high cost and complexity of deploying dedicated hardware, such as smart plugs for every device, or the requirement to purchase expensive, next-generation appliances with integrated IoT services. To overcome these barriers, Non-Intrusive Load Monitoring (NILM) has emerged as a transformative solution, enabling the disaggregation of total household power consumption into device specific insights using only the aggregated consumption retrieved by the main electric power meter. NILM algorithms have already been introduced in some existing works; nonetheless, the resulting infrastructures suffer from over-training biases (only data from the same house are used for both training and test sessions) and the lack of real-time responsiveness due to latency in current NILM systems. Therefore, this paper proposes an online IoT framework that bridges the gap between generalized energy models and the unique electrical signatures of individual homes. By leveraging a Deep ResNet architecture integrated with the CamAL algorithm, the system identifies appliances’ activation events. A core innovation of this approach is the use of a two-stage training strategy: a global model pre-trained on public datasets is refined through a localized partial-fit process using specific consumption data from the target residence. The performed tests demonstrated the usefulness and accessibility of the proposed solution, together with its higher accuracy with respect to existing solutions.