Hybrid Forecasting-Anomaly Detection Approach for Smart Building Energy Monitoring

Smart buildings rely on continuous monitoring and data-driven decision-making to optimize energy consumption, reduce operational costs, and detect faults early. In this context, energy forecasting enables proactive control and anomaly detection. This paper presents a hybrid machine-learning framework for short-term energy forecasting and anomaly detection using electrical measurements in innovative building environments. The proposed approach integrates real-world voltage and current data collected over 50 days to develop a real-time predictive model for active power consumption and identify abnormal behavior. The forecasting module is based on a Long Short-Term Memory (LSTM) neural network trained to predict active power using temporal and electrical features. The model achieved a Mean Squared Error (MSE) of 0.0007 and a Mean Absolute Error (MAE) of 0.012, demonstrating high accuracy in capturing short-term power fluctuations. A subsequent anomaly detection module based on an LSTM Autoencoder computes reconstruction errors to detect deviations from normal behavior. A total of 401 anomalies were detected based on reconstruction error thresholds, with strong agreement between forecast residuals and reconstruction-based anomalies. The results confirm the effectiveness of combining LSTM-based forecasting with unsupervised anomaly detection to enhance energy awareness and identify power quality issues, such as high harmonic distortion and low power factor. The proposed hybrid system enables interpretable diagnostics by pinpointing the root causes of anomalies, contributing to more intelligent, sustainable, and scalable building energy management solutions.