Data-Level Robust Federated Learning for Refrigerator Temperature Prediction with Anomalous Customers

Refrigerator temperature prediction using neural networks can improve both energy efficiency and customer experience in smart home environments. Training high-quality neural networks, however, requires large-scale data that accurately reflects diverse customer usage patterns. In practice, collecting and centralizing such data is challenging because of data privacy concerns. To address this issue, we adopt federated learning (FL), which enables collaborative model training across customers while preserving data privacy. In this study, we propose a robust FL that incorporates a data-level anomalous customer detection using an autoencoder model. This component is crucial because a subset of customers may exhibit anomalous data from various factors such as hardware malfunctions or abnormal usage patterns, which degrade the efficacy of the FL. The proposed framework, named DrFL, identifies customers with anomalous data and excludes them in the FL, thus preventing their adverse impact on the global model. Experimental results demonstrate that DrFL reduces prediction error by over 40% on average compared to existing FL methods, and maintaining robust generalization performance for new customers who did not participate in federated training. These results indicate that the proposed method effectively mitigates the influence of anomalous customer data.