Physics-Informed Neural Networks (PINNs) for Building Thermal Dynamics: Benchmarking on ASHRAE RP-1312 and BOPTEST

Zichao Li

Keywords: Physics-Informed Neural Networks, Building Thermal Modeling, HVAC Control, Energy Efficiency, Domain Adaptation, Noise Robustness, Edge AI

TL;DR: We present a hybrid physics-ML model for buildings achieves 19% energy savings via domain-invariant encoding and real-time (2.3ms) thermal dynamics prediction, validated across 6 benchmarks and real deployment.

Abstract: We present a Physics-Informed Neural Network (PINN) framework for building thermal dynamics that achieves cross-building generalization while maintaining physical consistency. Our model combines domain-invariant physics encoding with noise-robust training and computational graph optimization, outperforming five baselines across six benchmarks. Experimental results show 53\% lower RMSE than data-driven approaches, 3.4× better noise robustness, and real-time inference at 2.3ms/step. Deployment in real buildings demonstrated 19.3\% energy savings with sub-5\% comfort violations.

Submission Number: 16