The main infrsatructure of Pilot project 2 is located in Slovenia, particular at partners DIGITEH and ISKRA, while UL has its own Demo lab of factories of the future where developed approaches and results can be integrated, tested and evaluated.

• • 🏭 ISKRA, Ljubljanska cesta 24a, 4000 Kranj
    Specializes in physical systems and operational production lines. Offers real-time data collection via cloud servers, integrated solar panels and solar energy systems, cogeneration systems, and battery storage solutions.

    🖥️ DIGITEH, Tržaška cesta 315, 1000 Ljubljana
    Provides digital and virtual systems, including PC workstations and specialized hardware and software for developing digital twins of production systems and processes. Offers AI-powered digital twin solutions for sustainable and efficient energy management.

    🏫 University of Ljubljana, Faculty of Mechanical Engineering, Department of Manufacturing Systems and Processes, Aškerčeva 6, 1000 Ljubljana
    Hosts a demonstration lab of a Smart Factory, featuring digital and virtual systems, PC workstations, and various hardware and software tools for developing digital twins of production systems and processes. Includes AI-integrated digital twin solutions aimed at sustainable and efficient energy management.

1. Data and HMI 

A significant development since the previous milestone was the transition to a new data architecture. We have moved from direct database integration to using a new system based on REST API calls. These API calls are capable of retrieving real-time data, providing a scalable and flexible connection to the data source. However, to ensure stability during the current development phase, data retrieval is still performed only once per day. 

Alongside data acquisition, major progress has been made in the development of the Human-Machine Interface (HMI), which now consists of two integrated components: Admin HMI and User HMI. 

Admin HMI 

The Admin HMI is designed to configure the factory layout and associate machine data visually as shown in Figure 1. It allows administrators to set up new factory environment or change the already existing one. 

This ensures a flexible and scalable method for mapping any production or energy entity onto the layout, supporting consistent rendering in the user interface. To avoid duplication, each production line can be placed only once across all layouts. 

User HMI 

The User HMI provides a streamlined and interactive interface for monitoring the factory’s energy performance and production status. It is built dynamically based on the configuration created in the Admin HMI, ensuring accurate placement and labeling of machines and other entities. 

When opening the User HMI, users are first presented with several sections—such as Production Lines, Other Energy Producers, Energy Consumers and Overview, presented in Figure 2. Each section is tailored to the type of data and layout it represents. 

In the Production Lines section shown in Figure 3, users can load predefined factory layouts. These layouts display clickable buttons corresponding to each machine, positioned according to the configuration created in the Admin HMI. Clicking on a machine opens a popup that shows real and predicted energy consumption, production quantity, and additional metrics relevant to that production line. 

The Energy Producers section includes specialized popups depending on the producer type. For instance, photovoltaic systems display real versus predicted power output, along with environmental context indicators like sun or cloud icons. For batteries and Tesla units, additional reference lines indicate thresholds such as maximum capacity and 50% charge levels as seen in Figure 4. 

Each popup provides zoomable line charts that allow users to inspect historical trends and forecasts in greater detail. Navigation between sections is straightforward, and the interface ensures that only valid, non-duplicated entities are displayed per layout. 

The Overview section provides a high-level summary of the factory’s energy performance and upcoming management strategy. It contains two main views: Overview and Algorithms. 

In the Overview view, users can see the total energy used and produced for the current day, along with the percentage of consumed energy that is covered by internal production. This gives a quick insight into the factory’s energy self-sufficiency and efficiency. 

The Algorithms view displays the planned energy management strategy for the next 48 hours. This includes a visual representation of how different energy sources—such as grid, battery, photovoltaic, and cogeneration—will be used to meet expected consumption, based on forecasted data and optimization results. 

2. Anomalies in Energy Data 

With the transition to the new database and REST API-based data collection, several anomalies were detected in the incoming energy data. These included missing timestamps and abnormal spikes, often caused by temporary counter shutdowns or interruptions in data transmission. The data processing pipeline was updated to automatically detect and correct such inconsistencies, ensuring reliable input for the digital twin. These improvements help prevent distorted results in energy predictions and management strategies. 

3. Advances in Digital Model 

The digital twin model has seen significant technical enhancements, particularly in simulation continuity and intelligent decision-making. The simulation logic was upgraded to automatically detect and import only new data, allowing the system to resume from the last processed timestamp instead of restarting from the beginning. This improvement greatly enhances simulation speed and scalability. The research to transition from Technomatix Plant Simulation to Python based Digital twin for easier integration in different user systems was also done. 

Neural Network-Based Prediction 

The development of predictive models has followed a progressive path, moving from basic architectures to advanced ensemble methods. Initial experimentation began with stacked LSTM networks, which provided a solid baseline for capturing temporal patterns in energy data. Building on this, the team implemented an encoder-decoder LSTM architecture, enabling multi-step forecasting over a 48-hour horizon. Figure 6 presents results of comparison between predictions and real values for one machine in production line section. 

To improve accuracy and temporal alignment, attention mechanisms were introduced, allowing the model to focus on the most relevant time steps when generating forecasts. This was followed by the development of a Transformer-based model, which significantly improved the system’s ability to capture complex seasonal trends, sharp peaks, and irregular consumption patterns—especially across production shifts and energy generation variability. 

Finally, to leverage the strengths of each architecture, an ensemble model was implemented. This model combines the outputs of the LSTM, attention-based LSTM, and Transformer models using a meta-learner, resulting in more robust and stable predictions. The ensemble output is now used as the primary input for downstream energy management, offering both short-term accuracy and resilience to anomalies or rapid operational changes. The accuracy afforded by our ensemble model is up to 95%, meaning our predictions at best can deviate only 5% from real values – a very good result for timeseries predictions. The development of algorithms is shown as a block diagram in Figure 7. 

These predictive models are trained on real production and consumption data and updated regularly, ensuring the digital twin can anticipate factory behavior and enable proactive planning. 

Energy Management Optimization 

Building on the predictive forecasts, a dedicated energy management module determines the optimal strategy for using available energy sources across a 48-hour horizon. Two types of optimization algorithms were developed and tested: 

• The greedy algorithm, which evaluates energy decisions over short, rolling windows, prioritizing immediate benefits without considering long-term outcomes. 
• The more advanced smart algorithm, which performs full-period optimization across the entire 48-hour forecast, enabling globally optimal decisions based on future trends and constraints. 

These algorithms optimize not just energy flow, but also the sequence in which different energy sources are used—balancing grid power, battery storage, cogeneration, and renewables. The system supports two core objectives: 

• Cost minimization, which reduces total electricity expenditure. 
• Profit maximization, which aims to maximize net financial gain from energy production and export opportunities. 

Additional considerations include strategies for extending battery life, such as limiting deep discharge cycles and avoiding unnecessary charge fluctuations, which are factored directly into the optimization logic. 

The final optimized strategy is recalculated every 24 hours and visualized within the HMI, allowing users to monitor upcoming energy plans and better understand the rationale behind system behavior. This establishes a robust, closed-loop digital twin capable of predictive, constraint-aware, and goal-driven energy management. Figures 8, 9, 10 present partial HMI views of different energy management algorithms and their visualisation of energy sources usage. 

4. Conclusions and Next Steps 

Significant progress has been achieved in developing the digital twin, particularly in the areas of data integration, visualization, predictive modeling, and energy optimization. The transition to a real-time-capable database infrastructure has resolved prior data anomalies, ensuring clean and consistent input for simulation and forecasting. The Human-Machine Interfaces (HMI) for both users and administrators have matured into powerful tools for monitoring operations, visualizing predictions, and managing factory layouts. 

The evolution of predictive algorithms—from simple LSTMs to attention-based models, Transformers, and ultimately ensemble approaches—has resulted in reliable 48-hour forecasts for both energy consumption and production. These forecasts now serve as input for advanced energy management algorithms, which plan and optimize energy usage under real operational constraints. Both cost-minimizing and profit-maximizing strategies are supported, contributing to smarter, more efficient energy use. 

Next steps include: 

• Finalizing full integration of the digital twin into ISKRA Kranj and ISKRA Šebenik, with support for live data and daily updates. 
• Transitioning to Python based Digital Twin.