What question did this study set out to answer?

To validate a predictive collective control framework for large-scale kinetic facades using GNN and RL.

April 18, 2026Journal of Computational Design and Engineering0 citationsOpen Access

Predictive collective control of large-scale kinetic facades via Graph Neural Networks and Reinforcement Learning: From simulation to physical validation

SSSoochul ShinSeoul National University WLWonhyung LeeKorea Telecom (South Korea)YCYoungho ChaiChung-Ang University

Key Points

To validate a predictive collective control framework for large-scale kinetic facades using GNN and RL.
Developed a GNN-based surrogate model using 9,229 simulation samples.
Implemented a Proximal Policy Optimization (PPO) agent for real-time control.
Conducted summer-solstice stadium simulations to assess performance against an MLP-RL baseline.
Carried out on-site experiments to validate learned policies on physical hardware.
Achieved an 11.05% reduction in spectator heat load at peak solar conditions.
Increased field solar exposure by 20.63% at low-load periods.
On-site tests confirmed a 10.3% reduction in solar heat gain.
Decreased actuation effort by 25.4%, demonstrating effective energy efficiency.

Abstract

Abstract Real-time coordination of large-scale kinetic facades (KFs) is fundamentally challenged by the need to model strongly coupled module dynamics. This study presents the first on-site validation of a predictive collective control framework that integrates a Graph Neural Network (GNN)–based surrogate model with Reinforcement Learning (RL). Using a GNN surrogate trained on 9 229 high-fidelity simulation samples, a Proximal Policy Optimization (PPO) agent infers opening angles for adaptive solar control across thousands of interconnected units. Summer-solstice stadium simulations show that, compared with an independently controlled Multilayer Perceptron (MLP)–RL baseline, the proposed GNN–RL framework reduces spectator heat load by 11.05% at peak solar conditions (14:00) while increasing field solar exposure by 20.63% at low-load periods (16:00), reflecting a time-dependent Pareto trade-off between spectator comfort and turf daylighting. Crucially, the learned policy is deployed directly to physical hardware—without policy transfer—so a 1:30 mock-up reproduces emergent behaviors from global graph interactions. On-site experiments achieve a 10.3% reduction in spectator solar heat gain and a 25.4% decrease in actuation effort, demonstrating stable and energy-efficient real-time automation.

AI에게 질문

Bookmark

View Full Paper