A reproducible method to generate multi-building, multi-climate HVAC operation datasets with a stochastic exploratory...

Area of Science:

• Building energy systems
• Computational modeling and simulation

Background:

• Reproducible, controllable, and action-state-rich datasets are crucial for advancing building control research.
• Existing datasets often lack the diversity needed for comprehensive control strategy evaluation.

Purpose of the Study:

• To develop a method for generating multi-year HVAC operation time series across diverse buildings and climates.
• To expand control diversity through a stochastic supervisory controller.
• To facilitate transfer-learning evaluation and reproducible comparisons using distinct simulation domains.

Main Methods:

• A reproducible pipeline combining EnergyPlus (via Sinergym) and Modelica (via BOPTEST) was developed.
• A stochastic exploratory policy was implemented, introducing varied setpoint trajectories (drift, ramps, oscillations, jumps, noisy holds).
• Standardized 15-minute multivariate time series data (indoor temperature, weather, setpoints, HVAC power) were generated.

Main Results:

• The method successfully generated multi-year HVAC operation time series across heterogeneous buildings and climates.
• The stochastic supervisor broadened setpoint distributions beyond typical rule-based controller (RBC) schedules.
• Both datasets and reproduction code were released under FAIR principles.

Conclusions:

• The developed method provides a robust framework for generating diverse and reproducible HVAC datasets.
• The stochastic control approach enhances the richness of operational data for control research.
• The FAIR release of code and data promotes further research in building control, transfer learning, and robustness studies.