Research Info

Non-invasive load monitoring (NILM) is used to monitor household consumption at the appliance level, which is considered important for energy management and cost reduction. However, data confidentiality and privacy policies are viewed as major challenges in the training of NILM models. Federated learning (FL) is presented as a promising solution, as collaborative global training can be performed without raw data being exchanged. However, traditional FL aggregation methods such as Uniform averaging and its weighted forms are not capable of fully addressing user data heterogeneity, leading to weaker convergence and lower separation performance. In this paper, an adaptive method called Loss–Drift Federated Aggregation is presented, where dynamic weights are determined from three factors and assigned to each client: the amount of data, the local loss value, and the level of model drift. In this method, less weight is assigned to clients whose update behavior is observed to be unstable. Meanwhile, greater influence is given to clients whose behavior is more stable. A lightweight 1D-CNN model is used as the NILM predictor, making the approach suitable for deployment on edge devices. Experiments are performed on the REFIT dataset using a unified federated simulation setup consisting of 50 global rounds and 30 fine-tuning iterations. The results indicate that the proposed method yields consistent improvements over Uniform averaging and FedAvg-W, achieving up to 27% MAE reduction and showing greater robustness under heterogeneous conditions. These outcomes confirm the effectiveness of driftaware adaptive aggregation for privacy preserving NILM.