AN INTEGRATED DATA ENGINEERING AND MACHINE LEARNING FRAMEWORK FOR ENERGY PREDICTION AND FAULT DETECTION IN SMART BUILDINGS

Abstract

With the increasing demand for energy efficiency, this work proposes an advanced data-driven framework for energy prediction and fault detection in smart buildings. The system integrates multi-source data, including energy consumption, weather, and occupancy, to create a unified analytical pipeline. Advanced preprocessing techniques such as Z-score, IQR, STL, Rolling Z-score, and DBSCAN are applied for outlier detection, while linear and STL-based imputation ensure data continuity and reliability. For prediction, models including XGBoost, Random Forest, LSTM, Prophet, and MSTL are evaluated, where XGBoost achieves the highest accuracy and LSTM effectively captures temporal patterns. Fault detection is performed using Bollinger Bands, which identify anomalies based on deviations from expected energy trends. The proposed framework provides a robust, scalable, and efficient solution for intelligent energy management and proactive fault detection in modern smart building systems.