Data integrated Digital Twin to Enhance Building Energy Flexibility

Based on the work of Zahra Tabatabaei Mirhosseini who is the winner of the Netherlands (TVVL) student competition 2025.

Key words: Data integration, Digital Twin (DT), Model Predictive Control (MPC), Grid congestion, Demand side management (DSM)

Implementing advanced control for complex HVAC systems is challenging due to the need for accurate predictions under operational constraints. A data-integrated Digital Twin was developed and tested to forecast indoor conditions and energy consumption reliably, providing a model that enables predictive control to optimize winter heating peak loads.

Distribution grids have limited capacity and exceeding these limits can compromise stability leading to grid congestion. In the Netherlands, grid congestion is already occurring and is expected to continue, threatening renewable energy targets. Energy flexibility through demand-side management methods is an approach to mitigate grid congestion [1], which requires the effective adaptability of the current heating, ventilation, and air conditioning (HVAC) systems and building management systems (BMS).

This study addresses winter congestion issues in a Paris Proof building, in the Netherlands.  Success with a previous study on rule-based preheating strategy implemented in the building [2] led to exploring advanced solutions such as model predictive control (MPC). However, implementing MPC in practice remains challenging due to the complexity of HVAC systems, which consist of interconnected components such as sensors, actuators, and controllers. MPC adjusts setpoints to regulate heating while respecting operational constraints (e.g. physical limitations of devices), occupant comfort, and external disturbances like weather conditions. Achieving reliable predictions is therefore critical to prevent undesired outcomes; however, accurately estimating indoor temperature dynamics and energy consumption is challenging due to the strong coupling between these variables.

To address these challenges, a Digital Twin (DT) of the building’s heating system was developed with the capability of forecasting indoor conditions and energy use under various control strategies and weather scenarios. This predictive capability allows the DT to serve as the core model for an MPC framework, enabling the optimization of control trajectories to minimize energy consumption and peak demand while maintaining occupant comfort and satisfying operational constraints [3].

Methodology

The approach started with a rule-based preheating strategy to address the winter grid congestion, followed by model-based control using a Digital Twin within an MPC framework. A building management system (BMS) named InsiteSuite [4] handled data integration across hardware and software, enabling real-time monitoring, remote control, and reporting. A separate control platform named OpenToControl [5] provided smart logic, including predictive control, forecasting, and Digital Twin functionalities, ensuring efficient and flexible building operation.

Rule-based control

The building’s layout contains 4 different zones, as demonstrated in Figure 1, each zone including multiple rooms with separate temperature control. The rule-based control is intended for specific floors of the Purple zone (PS 1-4) and the Centrum zone (CM 2-10). Two preheating scenarios, manually implemented in the BMS, outlined in Table 1, are developed to shift the peak loads of the building to non-occupancy hours. The Yellow zone is working according to the normal schedule during the experiment. Therefore, this zone’s heating power data are normalized to match the preheated zone power consumption density, serving as a baseline for analysis.

Table 1. Heating flexibility scenarios.

Figure 1. The building’s layout. [6]

Model-based Predictive Control

Representing the physical system, the virtual part implements a closed-loop model in which the building’s thermal dynamics are regulated by a PI (C) controller. The system is subject to disturbances from weather conditions and occupant behaviour. Figure 2 shows the digital twin schematic of the HVAC system within the blue block, where Tsp (k), e (k), u (k), Hp (k), d (k), and Tin (k) are the temperature setpoint, the temperature error, the control action, the heat power, the disturbances, and the indoor temperature respectively for the time stamp k.

On the other hand, the building and the disturbances models are assumed as linear transfer functions G and H, considering Hp (k) as a saturated version (non-negative) of u (k). The indoor temperature can be represented as follows:

Where

,

Kp and Ki represent the proportional and the integral gains, and e (k) = Tsp (k) − Ti (k).

The MPC schematic shown in Figure 2 is presented as a service of the digital twin, where the real system is connected through InsiteSuite and the smart control logic through OpenToControl. Indoor temperature and disturbance measurements are collected at each time step k. The disturbances are used as inputs to a data-integrated forecasting service, which predicts their values over a prediction horizon of Np time steps using machine learning models. DT then uses these predicted disturbances to forecast indoor temperature and heating power over the same horizon. These predicted variables are used by the MPC to define system constraints and compute the cost function. Finally, an optimization module solves the control problem, providing optimal setpoint temperatures to both the digital twin and the physical system.

Figure 2. Close loop system block diagram for the HVAC (Digital Twin representation) and MPC scheme.

A simple mathematical model was used in the virtual part that describe the dynamic effect of the heat power and the disturbances as follows:

where α1, α2, α3, α4, and α5 are model parameters. Tout (k), O (k), R (k) and W (k) represent the outside temperature, the occupancy, the direct solar radiation, and the wind speed, respectively. The parameter set, including the controller gain kp, was estimated using the least squares method based on real data collected from 10 to 21 February 2025. For simplicity, the integral action of the controller is not considered in this study.

where J(Tin (k), Td (k), Hp (k)) is the cost function to be minimized, defined for energy as

and for peak power as

while maintaining the indoor comfort within the desired temperature Td (k), with ω1 and ω2 weighting the importance of each objective. The virtual model then simulates the indoor temperature and heating power based on predicted disturbances, providing inputs to the MPC to compute optimal control actions.

Results and Discussion

Rule-based control

Figure 3 demonstrates the peak power reduction across selected floors of the two CM & PS zones (Blue & Purple zones) compared to the Yellow zone. Accordingly, preheating in most cases causes reduction of the heating peak loads.

Figure 3. Peak load reduction of the selected floors through preheating.

The rule-based control determines when to start and stop preheating through manual inputs in the BMS. While this approach achieved promising reductions in peak load, it also increased the building’s overall power demand due to the higher preheating setpoints needed to offset heat losses. Therefore, the goal of improving efficiency and meeting the technical demands of complex BMS systems highlights the need for smart, model-based control to enhance optimization.

Model-based Predictive Control

The predictive controller simulation results are presented in Figure 4, which compares the indoor temperature setpoint, the actual indoor temperature, and the heating power under three different scenarios: (a) preheating (current case), and MPC strategies for (b) peak power reduction and (c) energy consumption optimization. Table 2 summarizes the performance results for the different control configurations.

Table 2. Summary of control results.

Conclusion

This study demonstrated the feasibility of integrating a digital twin with data-integrated forecasting and model predictive control (MPC) to optimize load shifting and reduce heating peak loads during winter congestion. Results show that MPC improves temperature tracking, lowers energy use, and smoothens peak thermal demand.

Digital twins and predictive control can overcome the limitations of rule-based control in complex BMS and HVAC systems, enhancing operational efficiency and indoor comfort. Integrating data and machine learning transforms rule-based control from manually determined settings into adaptive self-learning models. ML models learn dynamic patterns and handle complex systems and uncertainties, leading to improved energy flexibility strategies.

Acknowledgements

This project is funded by Topsector Energie Gebouwde Omgeving (TSE-GO) - TSEGO23003: https://buildinflexergy.nl/moprecef

References

[1]     E. de Winkel, Z. Lukszo, M. Neerincx, and R. Dobbe, “Adapting to limited grid capacity: Perceptions of injustice emerging from grid congestion in the Netherlands”, Energy Research & Social Science, Volume 122, 2025, 103962, ISSN 2214-6296, https://doi.org/10.1016/j.erss.2025.103962.

[2]     RCYP (REHVA Community of Young Professionals), Book of Papers 2025 (Milan, Italy), REHVA, 2025.  https://www.rehva.eu/fileadmin/user_upload/2024/Book_of_papers_2025_fv.pdf.

[3]     F. Tao and M. Zhang, "Digital Twin Shop-Floor: A New Shop-Floor Paradigm Towards Smart Manufacturing," in IEEE Access, vol. 5, pp. 20418-20427, 2017, https://doi.org/10.1109/ACCESS.2017.2756069.

[4]     InsiteSuite, “InsiteSuite: het merkonafhankelijke gebouwbeheersysteem”,n.d., https://insitesuite.nl/.

[5]     OpenToControl, "OpenToControl", n.d., https://opentocontrol.com/.

[6]     InsiteView, InsiteView(asr.nl), n.d.