TR2026-001

Quantile-SMPC for Grid-Interactive Buildings with Multivariate Temporal Fusion Transformers

    •  Hutchinson, S., Vinod, A.P., Germain, F.G., Di Cairano, S., Laughman, C.R., Chakrabarty, A., "Quantile-SMPC for Grid-Interactive Buildings with Multivariate Temporal Fusion Transformers", Advances in Neural Information Processing Systems (NeurIPS) Workshop on UrbanAI, December 2025.
      BibTeX TR2026-001 PDF
      • @inproceedings{Hutchinson2025dec,
      • author = {{{Hutchinson, Spencer and Vinod, Abraham P. and Germain, François G and Di Cairano, Stefano and Laughman, Christopher R. and Chakrabarty, Ankush}}},
      • title = {{{Quantile-SMPC for Grid-Interactive Buildings with Multivariate Temporal Fusion Transformers}}},
      • booktitle = {Advances in Neural Information Processing Systems (NeurIPS) Workshop on UrbanAI},
      • year = 2025,
      • month = dec,
      • url = {https://www.merl.com/publications/TR2026-001}
      • }
  • MERL Contacts:
  • Research Areas:

    Control, Machine Learning, Multi-Physical Modeling

Abstract:

Exogenous disturbances such as occupancy and weather strongly affect the per- formance of grid-interactive buildings, making accurate probabilistic forecasting essential for robust control. We propose a new MQF2-TFT architecture that captures both temporal and cross-variable dependencies for multi-horizon disturbance forecasting. To integrate forecasts into stochastic MPC, we develop a quantile- based sample method that enforces chance constraints despite non-Gaussianity. Experiments with real office building data demonstrates improved key performance indicators over baselines.