Machine learning-based surrogate optimization for multi-blind shading control in multi-occupancy open-plan offices

Efficient Model Predictive Control (MPC) for Venetian blinds in open-plan offices requires rapid optimization to respond to dynamic daylight and occupancy for minimize discomfort glare. Online surrogates is an ideal choice for optimizers to fast finding extrema of discomfort metrics. Unlike offline models, online surrogates are dynamically updated using small samples within each iteration step. This study investigated four online surrogate models—Gaussian Process (GP), Radial Basis Function (RBF), Extremely Randomized Trees (ERT), and Gradient Boosted Trees (GBT)—by comparing their optimization performance in maximizing multi-zone vertical eye illuminance within the discomfort threshold. Results show that RBF interpolation achieves an average convergence rate of 11.29, an extrema acquisition probability of 0.36, and a stability probability of 0.82, outperforming other methods in terms of convergence speed while attaining the highest extrema acquisition probability.