Building Energy Management Systems (BEMS) play a crucial role in optimizing energy efficiency, reducing operational costs, and enhancing sustainability in modern buildings, as they account for 40% of the overall energy consumption. However, the integration of diverse technologies, heterogeneous data sources, and varying communication protocols and mechanisms presents significant challenges, which restrict BEMS system adoption. In this paper, we explore these challenges, including interoperability issues, data standardization, cybersecurity/data privacy concerns, and scalability constraints. To tackle these challenges, we propose a co-design framework architecture that 1) facilitates seamless communication between different energy management services and 2) integrates components based on Internet of Things (IoT), Artificial Intelligence (AI), and cloud computing technologies. The framework also allows for addressing the challenges of BEMS systems, as well as provides a usercentered approach for increasing BEMS adoption, towards the development of energy-efficient smart buildings.
From the early start of the economy, credit risk has been study case for both researchers and financial experts. Financial credit scoring has been the most crucial process in the finance industry as it helps banks and financial institutions to make better decisions. The challenge arises in selecting the features that will make my decision better. This study proposes a feature selection approach using a genetic algorithm based on the information gain to improve classification performance. The genetic algorithm chooses feature subsets trough Logistic Regression. The validation is applied to the Home Credit Default Risk dataset. The experimental results demonstrates that the approach effectively selects features that improves the classification accuracy. The work suggests that evolutionary algorithms combined with mathematical measures can enhance credit risk prediction models and leads to better decision making in financial institutions.
In the domain of smart cities, Urban Data Platforms can enable the operation of digital services that rely on heterogeneous data sources, with the goal of enhancing the experience and well-being of citizens, as well as promoting sustainability and economic prosperity. In the present work, we demonstrate the Urban Data Management Platform, an orchestration of multiple software components that provides mechanisms related to the effective data and metadata management and storage, user authentication, while also aligning with interoperability standards. The demonstrated work provides a conceptual architecture of the platform and a technical implementation. In particular, we provide a high level architecture for the emerging urban data ecosystem, encompassing the links between urban data platforms, data spaces and urban digital twins. The implementation of the proposed Urban Data Management Platform leverages open source components, covering aspects related to data storage, user authentication and data space compliance. The resulting platform is tested as a core data enabler for urban geo-spatial services and digital twin applications in the context of the BUILDPSACE project. Future work on the platform will focus on interoperability and further alignment with data space principles.
Contemporary Ethernet networks rely on traffic classification techniques based on packet labeling fields such as the Type of Service (ToS). These mechanisms, while effective in prioritizing latency-sensitive flows, are vulnerable to manipulation. Malicious actors can falsely label low-priority traffic as highpriority (e.g., Voice or Video), thereby gaining unfair access to bandwidth and degrading network fairness. This study introduces a machine learning approach for anomaly detection, leveraging behavioral attributes of packets-such as size, interarrival time, and protocol port-to identify suspicious misclassifications. A comprehensive dataset was generated using the D-ITGBox traffic creation tool, emulating audio (G.721), video (H.264), background (ICMP), and best-effort (HTTP/Telnet) flows. Additionally, two anomalous datasets were synthesized by modifying the ToS fields of background and best-effort traffic to falsely represent them as voice and video, respectively. Three deep learning models-Convolutional Neural Networks (CNN), Multi-Layer Perceptrons (MLP), and Long Short-Term Memory (LSTM) networks-were trained and evaluated across 50 randomized iterations. Performance metrics including accuracy, precision, recall, and F1-score were analyzed. To validate performance differences, non-parametric statistical tests (Friedman and Wilcoxon) were applied. Results revealed that LSTM achieved superior recall and F1-score, indicating robust detection of sequential anomalies, while MLP excelled in precision. CNN demonstrated balanced but more variable performance. The findings advocate the application of deep learning models for Ethernet-based anomaly detection and set the groundwork for hybrid approaches and deployment in real-time environments.
Lockdown and home restriction have led to the overuse of video games. Video gaming refers to any activity involving games played on a personal computer, laptop, game console or any other type of device, either online or offline. This research protocol presents the outcomes of a study that evaluated the relationship between Covid-19-related home restrictions and gaming activity in Greece, during the period from March 23, 2020 to May 4, 2020. A total of 337 online users participated in the study (238 males and 99 females), with ages ranging from 19 to 50 years $(M=28.3)$. All participants were assessed using the Greek version of the IGD - 20 test, which was administrated online. Prior to the main test procedure, participants provided demographic information. The findings revealed that 16 participants (4.7%) were suffering from internet gaming addiction. Males $(n=112)$ spent more time playing games, compared to females. During the lockdown, there was a significant increase of more than 50%, in gaming hours among those who played more than 4 hours per day. Lockdown, quarantine, and residential restrictions seem to have contributed to the increase of gaming activity, with some individuals showing signs consistent with gaming disorder.
Deep learning models have been widely used to address financial trading challenges and develop profitable trading strategies. Recent studies have shown that incorporating sentiment information can enhance the performance of financial trading agents, as market sentiment extracted from news articles and social media often influences price movements. However, one major issue is the difficulty in distinguishing whether a positive sentiment value reflects a favorable past outcome or an expectation of good future results, and vice versa. This ambiguity can lead to suboptimal decision-making, as models may misinterpret sentiment trends and execute incorrect trading actions. To address this challenge, we introduce expert-guided features that simulate a real-life expert providing predictions on the future value of one or more financial indicators. The proposed model leverages this expert insight to assist the agent's decision-making process. By incorporating expert guidance, we conclude that our reinforcement learning-based trading agent can better navigate market fluctuations, selectively following expert advice when beneficial and disregarding it when misleading. Experimental results demonstrate the model's ability to differentiate between reliable and unreliable expert inputs, as well as its adaptability to various market conditions, leading to improved trading outcomes.
The increasing demand for electricity in modern households presents serious challenges for both environmental sustainability and grid reliability. Non-Intrusive Load Monitoring (NILM) has emerged as a practical solution for estimating individual appliance consumption using only aggregate power data. This paper proposes a novel NILM framework based on Denoising Autoencoders (DAEs), enhanced with a clusteringbased optimization technique that segments input sequences based on statistical features and trains separate sub-models per cluster. This approach is evaluated using minute-level smart meter data from Greek households, focusing on disaggregation performance for the dishwasher, washing machine, and washer dryer. Compared to the baseline DAE, the proposed model achieves substantial improvements, across all appliances, with the F1 score more than tripling in some cases, recall nearly doubling, and a 45% reduction in Root Mean Square Error (RMSE). The evaluation highlights the effectiveness of per-appliance clustering in enhancing disaggregation accuracy, especially for appliances with diverse and variable usage patterns. These results confirm the model's suitability for real-world smart home applications and energy management systems.
Nowadays, educational systems around the globe tend to integrate STEM in school programs in order to broaden students' perspectives and find alternative ways of integrating natural sciences into the curriculum. This paper presents an intelligent Internet of Things (IoT) parking system, developed using the micro:bit platform, AVR ATmega-328 microcontroller, and ESP-01 module. The proposed system incorporates a solar tracking mechanism to optimize energy harvesting through photovoltaic panels, thereby enabling autonomous battery charging. With the appropriate modifications, this system has the potential to become energy sufficient. Moreover, a ThingSpeak web page depicts data regarding power provided by the solar panel, voltage level on batteries, wind intensity and available parking slots. This system can be used in a simpler version for STEM education in schools; however, since it incorporates professional characteristics, it can be useful at the university level and, if scaled to real size, for commercial parking system applications.
The present research focus on the use of Named Entity Recognition (NER) for the anonymization of qualitative educational data in Greek. More specifically, it investigates the effectiveness of NER models for the Greek language in identifying sensitive information, such as tutor names, within open-ended student responses collected from course evaluation surveys at the Hellenic Open University (HOU). Five different NER models were examined, both individually and in combinations, to determine their performance in identifying proper names. The findings show that while all models demonstrate high precision and recall for non-entity classes, however significant differences are found in their ability to identify named entities. The Toolkit model is the best performing model, achieving high recall and precision, with some combinations of NER models improving recall but resulting in lower precision. The findings underscore the criticality of choosing appropriate NER models, especially in cases that involve category imbalance, and presents several trade-offs between precision and recall that one should consider in NER and anonymization tasks.
Large Language Models (LLMs) have been the focus of Artificial Intelligence (AI) research recently but evaluation of their performance demonstrated their limitations in various tasks requiring reasoning capabilities since responses of LLMs often contain erroneous answers and non existent facts. In this work we propose a solution to this problem by making use of Linked Open Data as a source of reliable information. Specifically, we propose an approach that leverages Large Language Models (LLM) in order to allow for automatic SPARQL query generation from natural language by either extracting classes and properties of the KG and match them with corresponding keywords in the user query or by providing example entries of the dataset to the LLM so that it can analyze its structure. Preliminary results demonstrate the potential of both methods.
In urban contexts the buildings' energy demand is quite significant, therefore their energy flexibility is important for successful Demand Response (DR) programs. Despite its promising features, blockchain adoption in DR poses significant security and privacy concerns, its long-term reliability being increasingly challenged by the rise of quantum computing, which threatens conventional cryptographic methods. Existing approaches do not adequately address this emerging vulnerability in the context of peer-to-peer (P2P) energy trading and buildinglevel flexibility aggregation. To bridge this gap, in this paper we propose the integration of a Quantum Random Number Generator (QRNG) into a blockchain platform, enhancing security and P2P energy flexibility transactions between residents and buildings. We show how quantum seeds can support the account generation and hashing of energy transactions in P2P energy trading. The Ethereum Keccak algorithm was adapted to incorporate a random quantum seed generated, reinforcing the blockchain's resilience against quantum computing threats and ensuring the integrity and immutability of energy transaction data. The evaluation results show quantum energy blockchain feasibility in managing energy flexibility transactions inside a building while featuring low additional overheads.
This study presents a machine learning (ML)-based approach for trip duration prediction using a large-scale mobility dataset from New York City (NYC) yellow taxi, incorporating both raw and engineered features to model spatiotemporal and fare-related factors that influence trip duration. Six regression models, including Linear Regression (LR), Support Vector Regressor (SVR), Random Forest (RF), Gradient Boosting (GB), XGBoost, and Multi-Layer Perceptron (MLP), were trained and evaluated using standard metrics, namely, Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Coefficient of Determination $(R^{2})$. The RF model achieved the best performance, with an MAE of 53.83 s, an RMSE of 143.38 s, and an $R^{2}$ score of 0.929. These results demonstrate the suitability of ensemble tree-based models for predictive analytics in intelligent transportation systems and outline future directions for enhancing performance using contextual and real-time data streams.
In recent years, Apache Spark has become one of the most widely used cluster computing frameworks globally, while FPGAs have increasingly been integrated into data centers for energy-efficient acceleration. In this paper, we explore the execution of computational tasks on FPGAs through the Apache Spark environment. To program the FPGA, we utilized the high-level synthesis platform SDSoC, where we developed a custom hardware platform named “zed.” Based on this platform, we implemented two applications: matrix multiplication and RGB-to-HSV image transformation, both written in high-level C++ code, with the goal of achieving faster execution through hardware acceleration. For the FPGA's operating system, we used PetaLinux 2017.4, specifically configured for the ARM Cortex ${ }^{\text {TM }}$-A9 processor. Additionally, we installed the Apache Spark environment on ZedBoard devices, creating a small-scale cluster consisting of two ZedBoards and a general-purpose computer for conducting our experiments. We evaluated the architecture using the two implemented applications. Experimental results show up to $11 \times$ speedup and $6.5 \times$ lower energy consumption for matrix multiplication, and up to $6.1 \times$ lower energy consumption for image conversion. Using two ZedBoards further improved execution time by $1.2 \times$. These results confirm the feasibility of executing Spark-based workloads on FPGA platforms and demonstrate the high performance-per-watt benefits of FPGA acceleration in distributed computing environments.
The rapid advancement of immersive technologies, such as virtual reality (VR), augmented reality (AR), and mixed reality (MR), has positioned immersive virtual learning environments (iVLEs) as powerful educational tools across many disciplines. Despite their widespread use and popularity, their effectiveness is often hindered by both educators' limited knowledge regarding their design, development, and application, and the limited knowledge of instructional design by game/interactive media developers. This paper details a graduate course whose aim was to introduce students to iVLEs and their design and development from an interdisciplinary perspective. The course followed a problem-based learning (PBL) approach that emphasized both the technical (e.g., game design) and instructional design aspects inherent in effective iVLEs. The paper begins with a detailed overview of the course followed by an overview of the course projects. Finally, some recommendations are provided to assist those wishing to implement and offer similar courses.
University students in Canada are experiencing a significant mental health crisis, with increasing rates of anxiety, depression, and stress. This crisis impacts academic performance and overall wellbeing. Traditional support services often face barriers such as stigma and limited accessibility. This work-in-progress paper presents the development of 'Dr. Calm,' an artificial intelligence- (AI-) driven chatbot designed to provide accessible and personalized mental health support to university students. 'Dr. Calm', an acronym for Cognitive Assistance for Life Management, employs a studentcentered design approach, integrating feedback from mental health professionals and AI experts. Utilizing ChatGPT, the chatbot aims to identify early signs of mental health concerns, offer evidence-based coping strategies, and track user progress. This paper outlines the project's objectives, design methodology, and anticipated outcomes.
This paper provides a brief overview of recent research in the field of Maritime Autonomous Surface Ships (MASS), with a focus on cybersecurity. A cyber-attack on the Nonlinear Autoregressive Moving Average (NARMA) controller of the MASS rudder is simulated, considering its effect on the rudder stability. The Kalman filter, as an additional device to the rudder controller, is used as a cyber-attack mitigation tool. The MATLAB simulation results provide insight into the system behavior without and with the Kalman filter, under the conditions of a simulated intrusion in the form of amplified noise imposed on the input and output signals of MASS.
Accurate classification of solitary pulmonary nodules (SPNs) as benign or malignant can be critical for lung cancer diagnosis and timely treatment. Traditional machine learning approaches rely on hand-crafted features for the classification task. This study explores the use of large language models (LLMs) for automated feature engineering to enhance SPN malignancy classification using a Random Forest classifier. A baseline dataset containing five standard radiological features was used to train a Random Forest model. Multiple LLMs, including GPT-4.0, Gemini, and others, were prompted to propose up to five new, clinically plausible features derived from or related to the original features. The suggested extra features were incorporated into new feature sets and evaluated using accuracy, sensitivity, and specificity metrics. All LLM-enhanced feature sets improved the classifier (88.52% accuracy), with the best results achieved using features proposed by GPT-4.0, reaching 94.64% accuracy, 96.16% sensitivity, and 93.54% specificity. Recurrent high-impact features included the SUVmax-to-Diameter Ratio, Margin Irregularity Index, and Nodule Growth Rate. LLMs show significant promise for automated feature engineering in clinical machine learning. Their ability to generate medically interpretable and performance-enhancing features can accelerate model development and improve diagnostic accuracy in lung cancer screening.
Accurate and efficient brain tumor classification from MRI scans remains a critical challenge in medical imaging, particularly given the variation in tumor types and the computational demands of deep learning models. In this work, we develop and evaluate a set of Convolutional Neural Network (CNN) architectures, referred to as BrainNet, BrainNeXt and BrainDil respectively, for the automated classification of glioma, meningioma, and pituitary tumors. Our models are based on the ResNet and ResNeXt families, with enhancements including additional residual blocks and the integration of dilated convolutions to capture multi-scale features with fewer parameters. We evaluate multiple configurations, including BrainDil with dilation rates of 2, 3, and 4, and BrainNeXt variants (50,101, and 152 layers). For additional comparison, these models were evaluated alongside traditional algorithms, including MobileNet and KNN. Across extensive experiments on two MRI datasets, BrainDil with a dilation rate of 4, achieves 97.10 % accuracy and a Kappa score of 0.9562, outperforming deeper and wider architectures in terms of both accuracy and consistency. These results highlight the effectiveness of lightweight yet expressive CNN variants for robust tumor-type-specific classification in clinical settings.
In the era of energy transition and smart technologies, the adoption and integration of innovative energy services-such as smart metering, demand response, renewable integration, electromobility, and energy trading platforms, among others-are critical for achieving sustainability goals. Such solutions often integrate novel technologies like blockchain, Artificial Intelligence, Machine Learning, and Internet of Things. This paper aims to develop a methodology for prioritising these smart energy services through multicriteria analysis, with the goal of better understanding priorities across different cases and different market actors. The methodology considers diverse criteria in five categories. These criteria are evaluated using the DEMATEL method. The weights of the criteria are used in order to prioritise the identified innovative energy solutions using the fuzzy VIKOR method. The results show that while Integrated Renewable Energy Systems are consistently prioritised across all expert groups, significant differences emerge regarding services like Virtual Power Plants and electromobility, highlighting the need for differentiated strategies based on stakeholder perspectives.
