Towards Open-Structure BMS: AI for Forecasting and Predictive Control

Key words: Building Management System (BMS), Machine Learning (ML), Model Predictive Control (MPC), Energy flexibility, Building automation

Europe’s energy transition redefines buildings as active grid participants. This study, based on advancements from an industry research project, introduces an AI-driven Building Management System that forecasts and optimizes energy use. Through predictive control and flexibility, it enhances grid stability, decarbonization, and efficiency, shaping the future of smart, responsive buildings.

Introduction

In the context of Europe’s accelerating energy transition, buildings are increasingly recognized not just as energy consumers, but also as active participants in providing flexibility to the energy grid. The ability of buildings to dynamically adjust their energy use is becoming a crucial asset for grid stability, decarbonization, and cost-efficiency.

To realize this potential, traditional Building Management Systems (BMS) must evolve. Current systems are often closed, vendor-specific, and reactive, lacking the AI-driven intelligence and openness needed to forecast conditions or optimally respond to dynamic grid signals. The Dutch industry research project Model Predictive Control for Energy Flexibility (MoPreCEF) [1] addresses this challenge by introducing an open-structured BMS with integrated data-driven advancements. Through forecasting and predictive optimization, this approach enhances building energy flexibility and paves the way for smarter, future-ready buildings.

This article outlines the core principles behind the open structure architecture, and the results from its pilot implementations, with a focus on what this means for the future of BMS design and the European building sector’s role in AI-driven energy flexibility.

Technical Framework

In the future, buildings will need smarter control systems that can anticipate energy needs. By generating and storing their own energy, and adapting to limited external supply, they can become more efficient and flexible. Studies suggest this could save over 20% of building energy use and cut peak electricity demand by up to 80%.

This should be managed not only by real-time data from traditional building-based climate control systems, but also by weather forecasts and predictions based on historical data from building installations, solar panels, electrical vehicles, and storage systems. This combination will enable better decisions and more effective system management.

Open structure BMS

The BMS of the future [2] will increasingly play a central role as a hub for various data streams within a building. To fulfill this role, a BMS must be able to provide the following basic functionalities:

·         Connecting various real-time data sources quickly and flexibly.

·         Connecting various historical data sources quickly and flexibly.

·         Storing historical data points and structures.

·         Calculating with real-time and historical data.

·         Clearly presenting real-time and historical data and related Key Performance Indicators (KPIs).

·         Making real-time and historical data available via open standards (API)[1].

And additionally:

·         Data-based predictions of building behaviour.

·         Supporting integration of decision algorithms for data-driven, predictive control (e.g., Model Predictive Control (MPC)).

·         Hosting predictive control systems that can perform real-time setpoint optimizations in connected systems (data sources).

To achieve an open structured system, it is important that the incoming and outgoing data streams are user adjustable. To make a BMS “open,” a generic Application Programming Interface (API) should be developed to enable data exchange with external sources and third parties. These connections must be standardized through a RESTful interface[2], which is linked on the backend to the specific data source’s API or protocol.

Forecasting and Predictive control

To enable predictive controls, modern BMS require new functionalities. One such functionality is an advanced prediction module (Figure 1), designed to integrate machine learning for model training and forecasting. Since training these models demands significant computing power and memory, the module should be supported by dedicated high-performance servers to handle these resource-intensive tasks. The module should be able to communicate through APIs with encrypted authentication over a secure connection.

The advanced prediction module for building management systems (BMS) is designed to communicate via an API over a secure connection for each user. Once authenticated, users can access a range of functionalities, such as:

Projects

The users can create projects for different applications and purposes. Within these projects they can store, modify and delete files and models.

Algorithms

The module provides a library of machine learning algorithms available for model training. Users can also view and adjust algorithm hyperparameters, following standard machine learning practices.

Models

Users can train models using their own data and save them within project folders. They can evaluate model performance, retrain if needed, and use the trained models for predictions. These predictions can inform control setpoints or other operational decisions within the BMS.

Data Files

Data for training and prediction is submitted to the module in standard formats (e.g., CSV).

The module is designed according to widely accepted machine learning principles, making it intuitive for users and easy to adopt. Additionally, a dedicated client library (e.g., for Python) can simplify communication with the module, enabling seamless integration into existing workflows and projects.

Figure 1. Prediction module with API based data communication.

Practical implementation

In this section, an example model predictive controller for smart electrical vehicle (EV) charging is shown with the above-described framework for data-driven prediction and control [3].

The implementation took place in a highly monitored office building in Breda, the Netherlands. This building has 4 charging stations (3-phase/ 16A / 11kW), PV panels with total capacity of 16.9 kW, an Air Handling Unit (AHU), a chiller and grid connectivity. The building has an average building load of 14kW.  The smart controller representation is shown in Figure 2.

Figure 2. Smart controller representation.

The goal of the MPC is to perform Demand Side Management by reducing power peaks and encouraging self-consumption of PV-generated energy, particularly in relation to EV charging. Currently, the building experiences morning power peaks that nearly reach the contracted transformer capacity, as shown in Figure 3.

Figure 3. Load/consumption profiles for a selected day.

The smart controller is set to use data from the building, EV users, and external services such as electricity price forecasts to optimize energy flows and minimize costs while maintaining user satisfaction (EV users) and operational efficiency.

The MPC, as shown in Figure 4, comprises separate submodules that interconnect yet can function independently. These are:

1.    The data preparation service

2.    The Optimization service

3.    The scheduler

In each optimization cycle, the scheduler triggers a sequence of tasks. At the start of each day, the electricity price forecast is retrieved from the grid energy provider. Subsequent tasks are executed every 15 minutes. The PV forecast is updated, and the building load is predicted by a machine learning model trained on historical and real-time BMS data. User input adds another key layer of information: EV users provide their vehicle’s current charging state and expected departure time, which are transmitted to the MPC via the BMS interface. All these inputs; PV generation, building demand, energy prices, and user schedules, are processed in the optimization module. Here, the MPC computes the optimal charging setpoints and current profiles that balance peak demand, cost, and self-consumption targets. Finally, these setpoints are sent through the BMS to the charging stations, ensuring dynamic, data-driven control that adapts continuously to real-time conditions.

Figure 4. MPC framework (how it operates).

The controller successfully reduces the peak charging demand and, consequently, the overall building peak demand by an average of 28% compared to an uncontrolled charging scenario. This reduction not only yields significant economic benefits by lowering maximum demand charges but also enables the potential expansion of EV charging infrastructure within the existing contracted transformer capacity. Depending on the specific energy tariff structure, average cost savings between 24% and 30% on peak demand can be observed. Importantly, these results were achieved without compromising the charging requirements or user satisfaction of EV drivers under real-world operational conditions.

Conclusion on BMS expansion for AI applications

Expanding BMS with AI enables predictive rather than reactive control, allowing buildings to anticipate demand and optimize performance. Effective use depends on integrating diverse data: sensors, forecasts, prices, and occupancy, into the BMS via standardized APIs and open protocols.

AI-driven BMS functions, in combination with mathematical optimization algorithms, demonstrate strong potential for enhancing energy flexibility, supporting demand response, and facilitating smart grid integration. This approach reduces energy consumption, lowers peak loads, and mitigates grid congestion, while also improving system resilience, scalability, and interoperability to advance both cost savings and long-term sustainability goals across building portfolios.

Despite the benefits, several challenges remain in expanding BMS for AI applications:

·         High Computational Requirements – Training and running AI models demand significant processing power and memory.

·         Data Quality and Availability – AI performance depends heavily on accurate, well-structured, and sufficient data from building systems and external sources.

·         Interoperability – Integration with diverse protocols and legacy systems can be complex.

·         Requirement for Expert Personnel – Developing, deploying, and maintaining solutions requires skilled data scientists, engineers, and domain experts, which may not always be readily available.

·         User Adoption – Facility managers and operators must adapt to new tools and processes, which requires training and change management.

Overcoming these challenges requires modular system design, sufficient computing resources, and intuitive interfaces to support broader industry uptake. Once integrated, AI applications can be scaled across subsystems such as Heating, Ventilation, and Airconditioning (HVAC), lighting, energy storage, and electric vehicle charging, further increasing their overall value in facility management.

Acknowledgements

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

References

[1]     "MoPreCEF," [Online]. Available: https://buildinflexergy.nl/moprecef.

[2]     InsiteSuite, [Online]. Available: https://insitesuite.nl/.

[3]     C. Lasitha , S. Walker, E. Petrova and P. Pauwels, "Towards portable model predictive control-based applications for demand side management in buildings," Energy and Buildings, vol. 347, no. 116257, 2025.