Abstract

The proliferation of IoT sensors in smart buildings has enabled extensive time-series data collection, providing valuable insights for predictive maintenance, operational efficiency, and energy management. This data can be used within an anomaly detection framework to automatically identify abnormal behaviors, enabling the detection of issues related to power consumption, control system failures, and sensor malfunctions. Additionally, it supports load forecasting, crucial for optimizing energy management by enabling efficient resource allocation, cost savings, and sustainable operations through accurate demand prediction. However, IoT devices are prone to issues like hardware malfunctions and transmission errors, which introduce data anomalies and complicate the effective use of collected data. A popular and effective anomaly detection framework trains autoencoders on minimizing the error between original and reconstructed sequences. By setting a threshold on the reconstruction error, abnormal sequences can be distinguished from the predominant regular patterns. However, this method is highly sensitive to architectural parameters and the nature of anomalies, making it difficult to develop a universally effective method or fine-tune a model without prior knowledge of the building and sensors settings. To address this, we propose a reinforcement learning-based neural architecture search approach to explore a manually defined search space and identify the optimal neural configuration through trial and error. While this method demonstrates competitive performance in discovering effective architectures that may be not intuitive, it assumes the availability of anomaly-free data for training, which is not practical in real-world scenarios. Hence, we also present an unsupervised feature bank-based model for anomaly detection in anomaly-contaminated time series. We integrate this with a recurrent denoising autoencoder, trained on data deemed anomaly-free, to replace identified anomalies with plausible patterns. This results in a combined anomaly detection and imputation pipeline that preprocesses data for downstream tasks, such as forecasting, in which we compare the performance of different models before and after preprocessing and demonstrate significant improvements. Despite these advancements, data scarcity remains a challenge in training effective load forecasting models for smart buildings, specially that concerns over the privacy of IoT sensor data make building owners reluctant to share their data, whether for transfer learning or for centralized model training, as it can reveal sensitive information about the occupants’ behaviors. To overcome this limitation, we introduce a federated load forecasting framework to exploit data from multiple buildings while maintaining data privacy. Acknowledging the diverse load profiles shaped by factors like size, location, and user behavior, we also investigate existing techniques to personalize the global model to accommodate each building’s specific characteristics.