BESTOpt: A Physics-Informed Neural Network Based Building Simulation, Control and Optimization Platform—A Case Study on Dynamic Model Evaluation

Zixin Jiang, Bing Dong

Keywords: Physics-informed machine learning, Building dynamic models, Energy optimization

TL;DR: BESTOpt, a physics-informed neural network platform for integrated building simulation and control, demonstrating great generalization over LSTM in unseen conditions.

Abstract: This paper presents BESTOpt, a modular simulation platform for building energy systems modeling. Unlike traditional tools that treat building dynamics, HVAC systems, and grid interactions in isolation, BESTOpt provides an integrated framework for dynamic modeling and control co-optimization. At its core is a physics-informed modularized neural network (PI-ModNN) that incorporates state-space-informed structural priors and hard physical constraints, enabling accurate, interpretable, and generalizable predictions of space air temperature. We evaluate the dynamic modeling module of BESTOpt against a purely data-driven baseline (LSTM) using synthetic datasets generated from EnergyPlus. While LSTM achieves lower prediction errors under normal conditions, BESTOpt demonstrates superior generalization in abnormal scenarios such as HVAC shutdowns, highlighting its effectiveness in control-oriented tasks where response fidelity is critical. The platform enables integrated building-to-grid applications at a large scale.

Submission Number: 30