Hierarchical data management and agentic intelligence in the CeDInt

Within the framework of the DIGEST project, we have completed the data acquisition phase for the CeDInt building at the Universidad Politécnica de Madrid. This marks a critical juncture in our effort to construct a semantically enriched, agent-enabled digital twin for intelligent building management. The CeDint building, already recognized for its advanced instrumentation and its role as a living lab for energy and smart systems research, provided an ideal foundation upon which to deploy our hierarchical management architecture. 

Our initial advantage stemmed from the maturity and robustness of the infrastructure established by the CeDInt research center. This facility is outfitted with a wide array of environmental sensors, energy meters, and control systems that support both passive monitoring and active control across diverse subsystems, including lighting, HVAC, and occupancy-driven automation. Importantly, this infrastructure is not only technically advanced but also operationally integrated, reflecting years of iterative optimization by the CeDInt team. Rather than replicating existing layers of hardware or basic data logging, our approach built directly on this platform, focusing on semantic integration, scalable orchestration, and intelligent interaction with building systems.

To realize this vision, we implemented a hierarchical data and device management strategy using ThingsBoard, an open-source IoT platform that supports scalable telemetry ingestion, rule-based processing, and API-based interaction. Unlike monolithic architectures, a hierarchical model allows for clear segmentation of responsibilities: at the base, raw data is collected from field-level sensors; in the intermediate layer, rules are enforced for alerting, filtering, and real-time visualization; and in the uppermost layer, services are decoupled from physical infrastructure through ThingsBoard’s RESTful APIs and message brokers (notably MQTT). This architecture aligns with recognized best practices in IoT system design (Bandyopadhyay & Sen, 2011), where modularity, interoperability, and resilience are prioritized.

A critical enabler of our future capabilities lies in the API infrastructure provided by ThingsBoard. Through this interface, we are now able to query real-time sensor values, retrieve historical data for offline analysis, and issue actuation commands through secure, authenticated channels. This allows us to move beyond the traditional use of buildings as passive data sources toward a more agentic model of intelligence, where digital agents operate within the building as autonomous, context-aware entities. These agents are envisioned not only as passive learners but as actors capable of closing feedback loops— predicting states, proposing or executing actions, and adapting to both environmental and user-driven changes.

The availability of structured, longitudinal data via API endpoints also enables the deployment of advanced machine learning techniques. From supervised learning models used to anticipate energy consumption and thermal dynamics, to unsupervised clustering techniques that help identify anomalous behaviors or novel usage patterns, the dataset collected from CeDInt now serves as a fertile ground for experimentation. Furthermore, reinforcement learning frameworks can be layered atop this infrastructure to develop adaptive control strategies, especially in areas such as HVAC setpoint optimization and load balancing, where explicit modeling is difficult and contextual feedback is necessary. The importance of such approaches has been widely recognized in recent literature (Zhao et al., 2021), particularly in the context of net-zero energy buildings and intelligent demand response.

Equally significant is the alignment of this system with the broader objectives of semantic interoperability, a core ambition of the DIGEST project. By maintaining not just the raw data, but also its contextual meaning—such as sensor location, calibration metadata, and device ontologies—our infrastructure contributes to the development of a semantic digital twin. This model serves not only current data analytics and control needs, but also
enables future scalability across buildings, districts, and potentially smart cities. The long term vision is to evolve from isolated data lakes to interconnected knowledge graphs where each entity, from a temperature sensor to a user interface, is semantically described and operationally accessible (Barnaghi et al., 2012).

In summary, the completion of the data acquisition infrastructure for the CeDInt building using a ThingsBoard-centered hierarchical architecture is not merely a technical achievement; it represents the strategic convergence of high-quality infrastructure, scalable management, and the methodological grounding necessary for developing agentic and learning-based systems. It paves the way for deep experimentation with AI techniques in real-world settings, all while ensuring that the system remains transparent, modular, and semantically rich. The groundwork is now laid for intelligent agents to interpret, act upon, and eventually co-manage the building environment in pursuit of both efficiency and user-centric outcomes.

References
Bandyopadhyay, D., & Sen, J. (2011).
Internet of things: Applications and challenges in technology and standardization. Computer Networks, 55(15), 4225–4238. https://doi.org/10.4236/cn.2011.31005

Zhao, H., Thilakarathna, S., Chhipi-Shrestha, G., Sadiq, R., & Hewage, K. (2021).
Predictive digital twin technologies for achieving net zero carbon buildings. Smart and Sustainable Built Environment, 11(3), 653–676.
https://doi.org/10.1016/j.sasbe.2021.12.003

Barnaghi, P., Wang, W., Henson, C., & Taylor, K. (2012).
Semantics for the Internet of Things: Early progress and back to the future. In 2012 IEEE International Conference on Semantic Computing (pp. 148–155). IEEE.
https://doi.org/10.1109/ICSC.2012.25