Figure 1. System Overview of the Centennial Campus at North Carolina State University

• Overview

As illustrated in Figure 1, a university campus functions as a city-scale energy ecosystem that integrates diverse building types, district heating and cooling plants, combined heat and power (CHP) systems, and increasingly, distributed energy resources (DERs) such as photovoltaics (PV) and battery energy storage. Because of this integrated and multi-layered infrastructure, university campuses provide a unique and strategically important testbed for Grid-Interactive Efficient Buildings and District Energy Systems (GEB & D). Unlike conventional residential or commercial districts, campuses exhibit highly dynamic and nonstationary energy demand patterns shaped by academic calendars, stochastic occupancy, and seasonal operational changes, all of which interact with weather variability and grid conditions. Although deep reinforcement learning (DRL) has attracted growing interest for building and microgrid control, existing studies have largely focused on single buildings or simplified systems, with limited attention to multi-asset coordination, resilience under extreme weather, and the substantial computational burden associated with training on high-fidelity physics-based models. To address these challenges, University Campus Flex proposes a holistic campus-scale digital twin that couples representative building clusters, district energy infrastructure, CHP systems, and DER aggregators while explicitly embedding stochastic occupancy dynamics and extreme-weather scenarios. By integrating surrogate modeling for computational acceleration and deploying an OpenAI Gym-compatible multi-agent reinforcement learning (MARL) environment, the framework enables cooperative agents to learn coordinated, resilient, and cost-effective control strategies under dynamic grid signals and disruption events such as outages. Ultimately, this project aims to deliver robust MARL control policies and an AI-ready research testbed that can transform university campuses from passive energy consumers into flexible, resilient, and community-supporting grid assets.

Figure 2. Conceptual diagram of a Grid-Interactive Efficient Building based on a high-fidelity digital twin

Figure 2 presents the conceptual framework of a commercial Grid-Interactive Efficient Building supported by a high-fidelity digital twin based on a calibrated simulation model. The objective is to develop advanced predictive control strategies for the HVAC operation of a multi-zone variable air volume (VAV) system, including approaches such as Model Predictive Control (MPC) and Reinforcement Learning (RL). By enabling load shifting under occupant comfort constraints, the proposed framework seeks to improve operational flexibility, reduce energy costs, and lower peak demand.