Sensor Fault Impact on Federated Learning for Electricity Demand Forecasting

Buildings contribute significantly to global energy consumption, necessitating advanced electricity forecasting methods to enable effective energy management and preserve data privacy. Federated learning (FL) has demonstrated promising performance in energy forecasting by protecting data privacy, yet its resilience against sensor faults remains underexplored. To address this gap, this study proposes an FL framework that integrates a lightweight model with an anomaly detection module to improve electricity forecasting accuracy in the presence of sensor faults. Within this framework, individual buildings preprocess hourly electricity data locally by detecting sensor faults (such as missing values and outliers) using statistical methods (e.g., Z-score and Median Absolute Deviation) and imputing missing data via seasonal-trend decomposition. The Time-Series Mixer (TSMixer) model, which employs multiple perceptron layers, is used to support resourceconstrained environments (e.g., smart meters). The proposed framework is evaluated using data from 30 educational buildings in the Building Genome Project 2. The results demonstrate that the framework consistently outperforms traditional approaches, where each building independently trains its own model using only local data. Moreover, the integration of the anomaly detection module yields a 38.2 % improvement in prediction accuracy compared to absence of the anomaly detection module. The framework maintains robust generalization with an 8.1 % performance degradation when applied to previously unseen buildings. These findings highlight the potential of FL-based forecasting systems to enhance energy management in real-world smart grid applications while ensuring data privacy and fault tolerance.