FRADS-GYM: A Reinforcement Learning Framework for Holistic Building Automation with Energy and Daylight Simulation

Title: FRADS-GYM: A Reinforcement Learning Framework for Holistic Building Automation with Energy and Daylight Simulation Abstract: Background: The optimization of building automation systems requires the simultaneous consideration of energetic and daylighting aspects to achieve both energy savings and high indoor comfort. Existing simulation environments for Reinforcement Learning (RL) in the building sector mostly focus on individual aspects and hinder the transfer of learned control strategies between different buildings. Objective: Development of a standardized RL framework that combines energetic and daylighting simulations and enables the systematic application of machine learning for individual building conditions (weather, location, typology, facade system, etc.). Methodology: FRADS-GYM combines EnergyPlus for energetic calculations with Radiance for daylight simulations, including glare assessment, in a unified Gym environment. The framework enables time-step-synchronous interaction between the agent and parallel simulations. Based on this, a modular system for reward functions was developed, enabling input scaling via adaptive standardizations and normalizations as well as unifying the combination of different optimization goals (energy, comfort, daylight). Unified interfaces facilitate transfer learning and the flexible exchange of building models. The Gym environment supports vectorization, allowing multiple independent copies of simulations to be executed in parallel to accelerate the learning process and hyperparameter optimization. Significance: While other Gym environments exist, FRADS-GYM uniquely combines holistic building simulation (EnergyPlus) with advanced daylighting (Radiance) and supports parallel execution. This enables reproducible and transferable research in building automation through standardized interfaces.