Optimization-Driven XGBoost-PINN Framework for Building Temperature Prediction

Building temperature prediction is critical for energy-efficient control in smart cities. We propose a novel hybrid framework that synergizes machine learning (ML) with operations research (OR) principles, combining XGBoost with physics-informed neural networks (PINNs) in a multi-stage optimization-driven approach. Starting from single-zone, single-day forecasts, we scale to multi-zone, multi-year predictions using Google’s Smart Building Simulator data. Our method optimizes physics-enhanced features, temporal encodings, and inter-zone interactions to mitigate uncertainty from noisy sensor data, achieving mean absolute errors (MAE) as low as 0.169°F for weekly multi-zone predictions. For long-term horizons, we employ OR-inspired ensemble strategies, maintaining robust performance up to 2.5 years. This work advances by enabling uncertainty-aware, energy-efficient building control for sustainable smart cities.