AI-Enabled Anomaly-Aware Occupancy Prediction in Grid-Interactive Efficient Buildings

This paper introduces artificial intelligence (AI)-enabled, anomaly-aware occupancy prediction methods to improve energy efficiency and demand response capability in grid-interactive efficient buildings (GEBs), while ensuring occupant comfort. First, the paper introduces an AI method based on Long Short-Term Memory Auto-Encoder (LSTM AE) for anomaly detection in time series data obtained from Internet of Things (IoT) sensors in GEBs. This AI-enabled method assists the building management system in early anomaly detection, thus reducing the risk of cascading failures. Isolating anomalous sensors and managing on-site resources based on accurate data can increase a building's reliability. Anomaly detection can also improve the accuracy of occupant activity prediction in GEBs. We use and compare LSTM and Graph Convolution Networks (GCN)-LSTM models to forecast occupant behavior on both raw data with anomalies and accurate data. Validation and assessment are conducted using data gathered from more than 1,100 IoT sensors located across a large academic building. The power consumption of the sensors is used as a metric for anomaly detection.