A Consumer Digital Twin for Energy Demand Prediction: Development and Implementation Under the SENDER Project (HORIZON 2020)

Abstract

This paper presents the development and implementation of a consumer Digital Twin (DT) for energy demand prediction under the SENDER (Sustainable Consumer Engagement and Demand Response) project, funded by HORIZON 2020. This project aims to engage consumers in the energy sector with innovative energy service applications to achieve proactive Demand Response (DR) and optimized usage of Renewable Energy Sources (RES). The proposed DT model is designed to digitally represent occupant behaviors and energy consumption patterns using Artificial Neural Networks (ANN), which enable continuous learning by processing real-time and historical data in different pilot sites and seasons. The DT development incorporates the International Energy Agency (IEA)—Energy in Buildings and Communities (EBC) Annex 66 and Drivers-Needs-Actions-Systems (DNAS) framework to standardize occupant behavior modeling. The research methodology consists of the following steps: (i) a mock-up simulation environment for three pilot sites was created, (ii) the DT was trained and calibrated using the artificial data from the previous step, and (iii) the DT model was validated with real data from the Alginet pilot site in Spain. Results showed a strong correlation between DT predictions and mock-up data, with a maximum deviation of ±2%. Finally, a set of selected Key Performance Indicators (KPIs) was defined and categorized in order to evaluate the system’s technical effectiveness.

1. Introduction

Digital Twins (DTs) play a significant role in energy demand prediction within modern energy systems by generating dynamic virtual representations of physical environments. By simulating various real conditions, energy consumption and supply can be forecasted more accurately, leading to higher energy efficiency.

The energy sector is currently facing complex challenges in terms of energy efficiency, sustainability, and ensuring a secure energy supply [1]. As both the demand for electricity and the need to reduce carbon emissions increase, it is vital to explore new approaches to improve the efficiency of existing energy infrastructure and to define a roadmap for future development [2]. Digital technologies can help in dealing with these issues, and the idea of DT is a very promising solution [3].

DTs, developed initially by Shafto et al. [4] to represent NASA’s equipment virtually, quickly became popular and useful. Nowadays, DT-driven decisions are used in many fields like manufacturing, logistics, healthcare, and energy [5]. A DT model uses data from sensors, machines, and other devices to create an accurate picture of the real world [6]. Moreover, technologies such as simulation, machine learning (ML), big data, cloud technology, and the Internet of Things (IoT) are used to build and improve DT models [7]. The combination of these technologies can help developers create complex DTs that copy the behavior of real systems and processes.

DTs play a significant role in energy demand prediction within modern energy systems by generating dynamic virtual representations of physical environments. The simulation of various real conditions can forecast energy consumption and supply more accurately, leading to increased energy efficiency [8].

Ghenai et al. [9] reviewed the application of DT technology across the entire energy value chain, from power generation and storage to energy use in buildings, transportation, and industrial processes. From this study, it became clear that the energy industry is increasingly interested in DTs, with the main goal of reducing energy consumption. Advances in ML and Artificial Intelligence (AI), together with more advanced control systems, will support the development of DT technologies. This will improve the efficiency and effectiveness of energy systems, encouraging the transition to clean energy. Bortolini et al. [10] reviewed the applications of DTs in enhancing building energy efficiency in four main areas: optimization design, occupant comfort, building operation and maintenance, and energy consumption simulation. They concluded that DTs in building energy efficiency are still an emerging field and that there is a growing interest in implementing DTs for building usage and maintenance.

Regarding the implementation of DTs to energy consumption, prior research has focused on residential applications. Meng et al. [11] introduced a new method for real-time temperature management in smart homes by combining DT and neural networks. The system continuously observed and evaluated the environmental information collected from units placed inside the house and then adjusted the temperature through an intelligent management system. Temperature changes were predicted by a DT, and a neural network model was used to improve the control methods to maintain thermal comfort. The results showed a noticeable improvement in temperature regulation due to the integrated system.

Fathy et al. [12] proposed a multi-layered DT for energy management. This system creates a “digital household” to reflect real-world energy consumption. By linking this “digital household” to a “digital power plant”, they developed an energy optimization model designed to increase energy efficiency specifically at the household level. The aim of the model was to make energy production more efficient by flattening daily energy demand. This was achieved by adjusting energy consumption patterns in homes to avoid peaks while meeting consumers’ needs and reducing their energy costs. The proposed system was tested using a two-year real-world IoT dataset from seventeen homes. The results showed that the model is effective in smoothing the overall energy demand and in reducing the average energy cost for each household.

Expanding this approach, Testasecca et al. [13] proposed a DT framework using Building Information Modeling (BIM), IoT, and ML to enhance energy efficiency and sustainability in areas like smart grids, microgrids, and district heating. Their research pointed out both the advantages and disadvantages of using DTs in Home Energy Management Systems (HEMSs), particularly mentioning issues with scalability, data reliability, and involving all relevant stakeholders.

Nakıp et al. [14] proposed a method to improve smart home renewable energy management. They used a neural network to predict energy trends, and they integrated the predicted data into scheduling to align production with consumption. Using this method, the accuracy of energy availability and demand forecasts was improved, leading to a more efficient use of Renewable Energy Sources (RES). The integration of forecasting into energy management, as also studied by Seyyedi et al. [15], improved energy efficiency and contributed to a more sustainable and balanced ecosystem in the smart home. Li et al. [16] combined artificial neural networks with the Gray Wolf Optimizer in a DT to improve solar microgrid management. The DT offered an advanced virtual model for real-time simulation and optimization, allowing the system to balance energy generation, consumption, and cost reduction, leading to an overall improvement of the microgrid’s efficiency and effectiveness.

DTs are revolutionizing home energy management, boosting efficiency, and promoting sustainability. These models address the complexities of modern energy systems by extending the use of IoT frameworks to home energy management. By creating real-time, virtual copies of physical systems, they enable real-time monitoring, predictive maintenance, and data-driven decision-making, which optimizes energy use and improves system performance [17]. Recent developments in ML and AI have further improved the performance of DTs, enabling intelligent reactions that adapt to user habits and environmental conditions [18].

De Lope et al. [19] applied the scope of AI-driven DTs in HEMSs in a wider area in order to improve hybrid renewable hydrogen systems for greater independence and performance. The success of their proposed model in optimizing real-time energy is clear from its use in a Spanish housing project. However, the complexity of the system makes it difficult to be flexible and expandable, while the dependence on central processing and complex ML raises concerns about security and the computing resources required for large-scale applications.

Cotti et al. [20] emphasized giving users control through a DT for smart homes based on ML. Their system uses unsupervised ML to model the way different device settings affect energy consumption. With an accessible web interface, users can set cause-and-effect rules and receive suggestions on how to reduce energy consumption.

Stogia et al. [21] presented a new system designed to encourage energy-saving habits in homes by combining IoT devices with DTs. They introduced a novel method that utilizes standardized 3D models with adjustable parameters, allowing simplified simulation and optimization of household energy settings. Their design significantly reduces installation complexity, improves adaptability, and provides users with up-to-date information on energy usage, indoor conditions, and practical advice on eco-friendly energy management. The findings demonstrate that the proposed system performs much better than conventional methods, achieving a 94% time savings in setup and a 98% reduction in memory requirements through the use of uniform parametric models and easy IoT device integration.

Combining prediction tools with energy management improved both the accuracy of demand forecasts and the quality of decision-making, resulting in more efficient power distribution across the microgrid [22]. These improvements allowed the system to better manage complex energy usage patterns, supporting more stable and sustainable solar microgrid operations. Seyyedi et al. [15] also demonstrated that integrating forecasting into home energy systems improved efficiency while helping balance consumption. Although the results are promising, further evaluation under varying conditions could clarify the broader potential of this method for RES.

DT technology is attracting growing interest from both researchers and businesses. Despite this, limited studies explore DTs for residential buildings, which could significantly improve residents’ daily lives. Therefore, this study focuses on creating a prototype DT for a home. The goal is to enhance the homeowner’s experience through methods beyond conventional Smart Home systems, such as automation based on simulations, predictive event analysis, and examination of historical and real-time data.

Bouchabou et al. [23] presented a method for addressing the problem of a lack of data that accurately represents the recognition of activities of daily living, using DTs to bridge the gap between training data and real-world data, enabling the creation of more diverse datasets. They proposed Virtual Smart Home, a simulator based on the Virtual Home simulator and designed specifically for simulating everyday activities in smart homes. To determine how realistic it is, they compared activity data recorded in a real smart apartment with a reproduced version in the VirtualSmartHome simulator. Furthermore, they showed that an activity recognition algorithm, once trained on data generated by the VirtualSmartHome simulator, can be effectively validated using real-world data collected in the field.

Gopinath et al. [24] introduced a redesign solution for Smart Homes using a DT. They conducted experiments on DT architecture and compared the results with typical IoT implementations. They concluded that DTs provide an added benefit to smart homes, especially since system design depends heavily on user preferences and situational assessment. Guizzardi and Fogli [25] presented the design and architecture of a DT for green smart homes. The system allows users to interact with their home environment through its DT, providing real-time visualization and control of home appliances, energy consumption, heating/cooling settings, lighting, and other parameters. The approach utilizes sensor data and predictive models to support optimal energy use while maintaining user comfort. They presented a pilot installation in the authors’ actual home and analyzed the way the use of DT can enhance environmental awareness and positively influence energy consumption. Finally, they discussed the advantages and challenges of this approach, such as the issue of privacy, scalability, and ease of adoption by end users.

Saadatifar et al. [26] investigated ways in which real-time temperature data from IoT devices affects people’s comfort levels regarding the temperature in their workplaces. They used an occupant-centric DT to provide users with a live picture of the temperature in different areas of the indoor environment, allowing employees to choose where to work based on their preferred temperature without sacrificing energy efficiency. Li et al. [27] developed a framework for building energy management using intelligent monitoring, utilizing DT. BIM, smart sensors, and IoT technologies enable the functions of four key components: the building’s physical space, a virtual twin space, a predictive control simulation engine, and twin big data. Following this standard framework and the function implementation methods, the DT can be effectively used for building operation and maintenance. The results indicated that incorporating digital technology in building intelligent control systems can achieve a maximum energy savings of 30%.

Clausen et al. [28] designed and implemented a DT framework for buildings, where the climate-controlled spaces are represented digitally. Within this framework, DTs are parametric models incorporated into a general control algorithm, which utilize weather forecasts, current and anticipated occupancy data, and the actual state of the controlled environment to perform Model Predictive Control. They showcased the DT framework’s application through a case study at the University of Southern Denmark, where a DT manages heating and ventilation systems, and demonstrated that their system can achieve comfort levels comparable to those of existing commercial building management systems while also allowing for the implementation of strategies aimed at boosting energy efficiency. Tagliabue et al. [29] introduced a framework for transitioning from traditional, static sustainability assessments to a dynamic method leveraging DTs and the IoT by facilitating real-time assessment and management of diverse sustainability factors from a user-focused perspective. To validate the framework, it was tested on a pilot building at the University of Brescia, using several example applications. This educational facility supports the daily routines of engineering students by continuously interacting with sensor-equipped systems that monitor indoor comfort, air quality, and the building’s energy consumption, allowing the optimization of the balance with renewable energy generation.

Corssac et al. [30] demonstrated two applications, illustrating how a two-way connection between a residence and its digital representation can offer homeowners analysis and simulation-driven automation. The initial use case enables users to view the historical and current status of their home appliances regarding energy consumption. The second use case, relying on heating simulations, identifies the optimal time to activate a heater to improve thermal comfort while minimizing energy usage.

Sayed et al. [31] introduced a new method for improving energy efficiency in homes by integrating a DT on the Home-Assistant platform, which allows homeowners to create and manage virtual versions of their homes using various IoT devices and sensors. They developed a system that provides real-time energy consumption data, energy-saving suggestions, and a data-driven approach to detect occupancy. The proposed system was tested in a small-scale pilot study, which showed a significant 80% positive user rating, demonstrating its success in offering helpful energy-saving tips. The DT recorded the user’s activity, their engagement with smart devices, and the lab’s ambient environment. The findings suggest significant opportunities to lower energy consumption and expenses while ensuring user comfort.

DT technology plays a key role in the development of smart lighting systems for smart homes, proposing models that improve stability, adaptability, and user experience while addressing challenges such as standardization, interoperability, and development costs [32,33].

Arsecularatne et al. [17] explored the use of DT to better manage building energy and understand occupant behavior. They evaluated the extent to which DTs optimize energy usage through a detailed literature review combined with a scientometric analysis. Key challenges identified include problems with interoperability, concerns about privacy and data quality, and the need for better integration between digital and real-world interactions. Their findings emphasized the importance of standardized frameworks to guide the implementation of DTs and highlighted areas needing more research, particularly improving cybersecurity and including occupant behavior in DT models. Recently, Arowoiya et al. [34] reviewed current research on DTs for thermal comfort and energy efficiency, and they concluded that further investigation is needed into extended reality, indoor air quality, and personalized thermal comfort models. Analyzing the methods, technologies, algorithms, and approaches used in DTs revealed that the use of sensors for environmental measurements is restricted, mainly focusing on air temperature and relative humidity. They recommend that future research should focus on occupant feedback, specifically on the need to better integrate their behavior and personalized comfort.

The SENDER (Sustainable Consumer Engagement and Demand Response) project [35] aims at the active participation of consumers in the energy sector with innovative energy service applications to achieve proactive Demand Response (DR) according to the needs of the target population, optimizing at the same time the usage of RES. The current paper shows the development and implementation of a consumer DT developed to represent occupant behaviors and energy consumption patterns using Artificial Neural Networks (ANN).

2. Development and Implementation of the Digital Twin

According to ISO/IEC WD 30173 Digital twin—concepts and terminology [36], a DT is a digital representation of an entity that offers an integrated view across its lifecycle. It synchronizes with a physical entity using data connections to enhance performance. ISO/IEC WD 30172 Digital twin—use cases [37] presents use cases across various domains, using a standardized template developed by comparing existing templates from IEC 62559-2 [38], ISO/IEC TR 22417 [39], JTC 1/AG 8 [40], and SC 41/AG 25 [41]. The interactions between external actors and the system to attain certain goals are defined by IoT and Digital Twin Personal Wellbeing Index (PWI) guidelines. The use case template includes two parts: description and drawings/diagrams. The description part covers the use case name, application area, scope, objectives, digital entities, actors, lifecycle, Key Performance Indicators (KPIs), digital infrastructure, and more. The drawings/diagrams part includes use-case drawings, data flow, sequence, and deployment diagrams. Each use case defines system functionality through interactions between users (actors), the system, and their goals.

The DT is designed to create a digital representation of a building, taking into account the occupancy activities inside it. Within the SENDER project, consumers are involved in both energy consumption and production, aiming to become prosumers. Initially, consumers’ behavior inside a building or a residence is monitored to define and model energy patterns, translating them into usable data to be simulated and to generate energy forecasts. Following this, the simulated outcomes are analyzed and compared with real-time data, and they are refined to more accurately reflect actual data, reducing the discrepancy between predicted and actual energy consumption. This precise simulation will contribute to optimizing the balance between occupant comfort and energy usage.

To meet the demands of the SENDER solution, an ANN enhancement was chosen for the consumer DT, which extends beyond the scope of ML in AI, moving towards Deep Learning (Figure 1), to create a more precise DT of occupant behavior than typical statistical methods [42]. This strategy is based on PARAGON’s [43] human passenger DT development from the aeronautics sector’s EC project, HEACE (Health Effects of Aircraft Cabin Environment) [44], adapting technology from a research area generally more advanced than other fields. In the HEACE project, a similar DT was developed to model the behavior and reactions of passengers on long-distance flights. The key benefit of the AI component is its capacity for continuous learning, leading to high individualization and accuracy.

DT, based on occupant behavior, generates profiles of occupants’ actions using monitored data. This provides a combined perspective of consumers for grid operators or aggregators, enabling them to identify, exploit, and commercialize flexibility opportunities. The integrated ANN learns through examples. Data collected from each consumer serves as training material for the AI, which is adapted to replicate observed behaviors as accurately as possible, even inferring finer details when necessary. The ANN starts with basic statistical behavioral patterns, and then it is trained using a subset of the measurement data. After training and operational validation with more data, the ANN continues learning from real-world examples, refining its performance to better reflect the occupant’s behaviors. To simplify DT design and maximize its performance, ANN is designed not only to be adapted to individual occupants but also to segment each occupancy behavior by season (spring, summer, autumn, winter), creating a seasonal sub-layer for every occupant, allowing for more precise and direct adjustments for each season of data.

The International Energy Agency (IEA)—Energy in Buildings and Communities (EBC) Annex 66 “Definition and Simulation of Occupant Behavior in Buildings” [18] and the Drivers-Needs-Actions-Systems (DNAS) framework ontology for occupant behavior standardization were included in DT development, and the interfacing was ensured by the Functional Mock-up Interface (FMI) standard. The DNAS Framework is an energy-focused ontology designed to standardize and model occupant behavior for integration into various energy interfaces and simulations. Rather than generating a completely new dataset from occupant behavior data, the DT follows the DNAS framework’s standards and methodologies. Obviously, some modifications and adjustments to the DNAS framework are necessary to align with the SENDER project’s requirements. This approach also promotes sustainability in data utilization, conserving resources by removing unnecessary components and grouping similar behavioral data. Figure 2 illustrates the DNAS Framework’s implementation with the DT library.

For the development of the initial DT version, a mock-up for simulations was created. This mock-up environment is a fully operational testbed, which serves as the DT’s workspace during the development phase. Artificial data was used to represent buildings, their occupants, and occupancy behaviors, reflecting the expected occupancy data at the three pilot sites of the SENDER project. Initially, the ANN component of the DT was trained using a subset of the mock-up data. The remaining subsets were then used to validate the ANN component and simulate the continuous learning process expected with the SENDER solution, as it operates at the pilot sites.

To validate this DT application, the open-source Building Energy Management System, EnergyPlus (v.22.1.0) [45], was employed as a co-simulation manager using the FMI. The simulation covered the entire year of 2022, 8760 h in total. The test site chosen was Alginet (Valencia, Spain), as the main goal was to validate the DT against the expected real-world data. The baseline for performance testing was based on a statistical method typically applied in large professional sites [42]. To ensure compatibility and simplicity, the mock-up building was a multi-story office with a total area of 1440 m², occupied by office employees working in three different shifts per day. Climate data for Alginet was retrieved from “Climate Analytics” [46].

Figure 3 depicts the comparison between the mock-up occupancy data and the corresponding statistical prediction based on Annex 66 guidelines [36]. The main issue highlighted is that the Annex 66 statistical method relies on shift averages, which are reliable when results are averaged over a week or longer. However, this method does not meet the needs of the SENDER project, which requires measurements in multiple-minute subsets within an hour, rather than multiple-hour subsets within a day. For SENDER, accurate predictions within a few minutes are crucial for the overall functionality of the SENDER solution.

Following, Figure 4 presents the same mock-up occupancy rate compared to the corresponding DT prediction. The graph shows a strong correlation between the DT prediction and the mock-up data during a simulated day, with a maximum deviation of ±2%. Consequently, the DT implementation, which includes the AI component, has demonstrated performance that aligns with the requirements of the SENDER project.

In summary, the ANN methodology can be described as follows:

(i)

A deep learning–based ANN enhancement of the consumer DT, designed to generate occupant behavior profiles and incorporating seasonal sub-layers (spring, summer, autumn, winter) for each occupant.

(ii)

Data-driven training through examples, where monitored consumer data serve as training material and the ANN is initialized from basic statistical behavioral patterns.

(iii)

An initial training and operational validation phase using mock-up subsets of measurement data, followed by continuous learning from real-world examples during operation.

(iv)

Methodological alignment and performance comparison with the IEA EBC Annex 66 guidelines and the DNAS framework for occupant behavior representation.

A direct, simple approach was adopted using a backpropagation algorithm, with a 3–4 layer ANN architecture. Groupings for input and output parameters were defined according to Annex 66 recommendations. The aim was to increase predictability to a finer time resolution, i.e., in minutes instead of hours, while keeping the architecture simple and robust, suitable for industrial application scenarios. Evaluation metrics were defined to assess minute-level predictions of occupancy and consumption data. An acceptable limit of ±2% deviation from the corresponding actual data was set as a desired level of accuracy for reliable DT performance.

Access to real-world measurement data is challenging, and available empirical datasets are limited. In terms of the technical part of the SENDER project, the development and implementation of a consumer DT for energy demand prediction was crucial. To achieve this, real-time and historical data were collected and processed across three different pilot sites. Three European countries—Austria, Spain, and Finland—are the pilot sites for the SENDER project, with an initial goal of collecting one year of measurement data from approximately 400 households. Due to external constraints, including supply chain disruptions caused by the COVID-19 pandemic and the Evergreen ship incident, which affected equipment availability and installation schedules across pilot sites, the final number of operational households with sufficiently reliable data was reduced, covering a minimum monitoring period of six months.

Although this is a deviation from the original goal, the SENDER project provided the most extensive dataset among the sister projects HESTIA [47], ACCEPT [48], iFLEX [49], and REDREAM [50]. The availability of this data supported the development of the proposed DT architecture, despite limitations in data homogeneity and overall duration. The present work focuses on the methodological formulation and architectural design of the digital twin, without addressing detailed quantitative performance analysis. For completeness, a standard split of the available measurements was applied, with 90% allocated to training and 10% reserved for validation during the ANN configuration phase. Figure 5 presents the methodological workflow for the development of the Consumer DT for the SENDER Project.

3. Key Performance Indicators Selection

Key Performance Indicators (KPIs) are essential for the evaluation of the energy system’s performance. The selected KPIs, their category, and their priority are presented in

Table 1. Selected KPIs and their priority.

KPI ID	Category	KPI Name	Priority
1	Grid Performance and Energy Efficiency	Level of losses in low-voltage networks
2		Voltage quality performance of electricity 1
3		Percentage utilization of electricity grid elements 2
4		Reduction in energy losses
5		Reduction in peak demand
6	Consumer Energy Behavior and Savings	Energy demand and consumption
7		Energy savings
8	System Flexibility and Market Participation	Increased system flexibility from energy players
9	Computational Performance	Amount of data transferred
10		Simulation speed for real-time applications
11		Data storage capacity
12		Processing speed
13	Operability and Feasibility	Operability
14		Feasibility of the asset configuration
15		Bandwidth capacity
16		Operability of the interfaces
17		Robustness of the interfaces
18		Feasibility of the interfaces’ configuration

¹ Voltage variations. ² Lines and transformers.

. A traffic light color scheme is employed for illustrative purposes: high priority is indicated in green, medium priority in yellow, and lower priority in red.

The KPIs were organized into five categories to provide a clear view of DT capabilities. The first category, “Grid Performance and Energy Efficiency”, includes indicators that measure the DT’s technical impact on the grid, such as losses in low-voltage networks, voltage quality, utilization of grid elements, reduction in energy losses, and peak demand reduction. The second category, “Consumer Energy Behavior and Savings”, focuses on the way that DT captures occupant energy patterns, including energy demand, consumption, and potential savings. The “System Flexibility and Market Participation” category evaluates the DT’s role in supporting DR actions among multiple energy stakeholders. The “Computational Performance” category covers aspects such as data transfer, simulation speed, storage, and processing speed. Finally, “Operability and Feasibility” assesses how practical, reliable, and robust the DT is, including interface performance, system configuration, and network capacity. The KPIs are primarily intended to evaluate the performance of the consumer DT. These include KPIs related to consumer energy behavior (6–7), computational performance (9–12), and operability and feasibility (13–18). Additional KPIs, such as grid performance, energy efficiency, and system flexibility (1–5, 8), provide insights into the broader SENDER ecosystem.

4. Conclusions

The consumer DT developed within the SENDER project creates a digital avatar of buildings, monitoring occupancy behaviors in order to optimize energy demand and consumption, by balancing environmental comfort inside a building with energy efficiency. As part of its development, the DT uses ANN to better model and predict energy demand patterns. IEA Annex 66 and DNAS ontology were applied in a similar DT already developed for aviation in order to achieve a successful DT forecast for other seasons or energy usage patterns.

In terms of this study, mock-up data and a full-year simulation using EnergyPlus for a large building in Alginet, Spain, were compared to show the correlation between real-world and virtual occupancy patterns. Comparisons showed a maximum deviation of ±2% in occupancy prediction, confirming the validity of the model.

The DT’s ability to learn and adapt continuously, combined with its integration into a co-simulation environment via FMI, helps grid operators better understand flexibility and manage energy more effectively. The selection of comprehensive KPIs further ensures that the system can be evaluated and improved.

This work represents the first step in DT development within the SENDER project. In the next phase, the developed DT will be used to analyze real-world data collected from pilot sites in Austria, Spain, and Finland, further validating its applicability across different regions.