"Real-Building Control via Model-based Inverse Reinforcement Learning"

"This dataset accompanies the paper REAL: Real-world and Energy-efficient Adaptive Inverse Reinforcement Learning for Smart Buildings and contains real-building operational logs collected from a deployed Variable Air Volume (VAV) HVAC system under three different control strategies: (1) a conventional Rule-Based Controller (RBC) used as the operational baseline, (2) a state-of-the-art model-free Inverse Reinforcement Learning (IRL) controller, and (3) the proposed REAL controller, a model-based IRL framework designed for adaptive and energy-efficient building control.The dataset includes synchronized time-series measurements of both system states and control actions collected from multiple real building zones. State variables include zone air temperature, outdoor air temperature, and time-of-day information, while action variables include VAV damper position and airflow rate. The logs were collected at fixed control intervals during real operational deployments in a university building environment.This dataset is intended to support research in intelligent building control, reinforcement learning for cyber-physical systems, inverse reinforcement learning, energy-efficient HVAC operation, and data-driven control benchmarking. It enables reproducible evaluation and comparative analysis of traditional rule-based control, model-free learning approaches, and hybrid model-based IRL methods in real-world building environments.The dataset may be used for tasks such as controller benchmarking, offline reinforcement learning, reward modeling, system identification, occupancy-aware control studies, and building energy analytics. By releasing these real-building operational traces, this work aims to facilitate reproducible research and accelerate the development of practical AI-driven HVAC control strategies for smart buildings."