Publications

Prompting fairness: Learning prompts for debiasing large language models

Authors: Camelia Lemnaru, Cristian Andrei Rad

Large language models are prone to internalize social biases due to the characteristics of the data used for their self-supervised training scheme. Considering their recent emergence and wide availability to the general public, it is mandatory to identify and alleviate these biases to avoid perpetuating stereotypes towards underrepresented groups. We present a novel prompt-tuning method for reducing biases in encoder models such as BERT or RoBERTa. Unlike other methods, we only train a small set of additional reusable token embeddings that can be concatenated to any input sequence to reduce bias in the outputs. We particularize this method to gender bias by providing a set of templates used for training the prompts. Evaluations on two benchmarks show that our method is on par with the state of the art while having a limited impact on language modeling ability

A Hybrid Machine Learning–Genetic Algorithm for Optimizing Surface-Mount Technology Planning

Authors: Adrian Petru Groza

We tackle the problem of improving the Surface- Mount Technology (SMT) process planning in an automotive manufacturing setting. Current simulations show low accu- racy across production lines as the existing approach relies on predefined setups rather than adapting to product-specific configurations. We propose a hybrid framework that couples machine learning with a genetic algorithm to generate product- specific plans. Our solution involves three tasks: (i) assigning boards to lines, (ii) allocating components to Pick-and-Place (PnP) machines, and (iii) balancing workloads across machines. Our hybrid pipeline embeds supervised learning in a genetic optimizer. A multi-class classifier selects feasible PnP head con- figurations per Bill of Materials (BOM) part number (precision = 0.73). A genetic algorithm assigns components to compatible feeder tables/machines, while a regression model estimates table cycle times (R² = 0.88). The fitness jointly optimizes Components Placed per Hour (CPH) and Line Balancing (LB) under process constraints. Different mutation methods are explored, revealing that mutation based on balancing the workload by leveling the number of placements on the tables with minimum and maximum cycle time results in an LB of 0.83, with a CPH of 0.37 and an average delta cycle time of -3.27% across 105-part numbers

ConvU-NExT: An Asymmetrical Encoder–Decoder for Denoising Low Dose CT

Authors: Adrian Petru Groza

Low-dose computed tomography (LDCT) is a medical imaging modality designed to minimize ionizing radiation exposure while maintaining the ability to produce detailed cross-sectional images. It is particularly valuable in scenarios requiring repeated imaging, such as cancer screening, follow-up examinations or pediatric diagnostics, where reducing radiation dose is critical to patientsafety. For example, to reduce noise by half, fourtimesthe radiation dose isrequired in the slice. The goal isto achieve postprocessed LDCT images with comparable quality to those obtained from standard-dose CT imaging. We start with a brief overview of the CT procedures and their limitations. Then we introduce a novel denoising method based on an asymmetric integration of the ConvNeXt backbone with the U-Net architecture. This novel approach obtained 2–3 times less noise than the original LDCT, having a 10%–20% increase in performance compared to U-Net implementation, checked against three metrics MSE, SSIMLoss and combinations of both. The results suggest that: (i) augmenting the images with specific noise, obtained from water phantom CT scan test, while training yieldssuperiorresults compared to generic noise augmentations; (ii) a larger kernelsize better extracts features and (iii) a smaller kernel size was mandatory for feature reconstruction

AlloyGraph: Data and Evaluation Results for Multi-Agent AI Superalloy Property Prediction

Authors: Alexandru Lecu, Adrian Petru Groza

Training data (77 alloys from the Nickel Institute handbook), evaluation data (88 alloys from manufacturer datasheets), prediction results for six model configurations, chatbot evaluation benchmarks (250 MCQ questions, 100 RAGAS questions, 12 expert-graded questions), inverse design results (20 target specifications), and OWL ontology for the AlloyGraph platform. Associated repository: https://github.com/AlexLecu/AlloyGraph

OCTA-Based Biomarker Characterization in nAMD

Authors: Adrian Petru Groza

We aim to enhance ophthalmologists' decision-making when diagnosing the Neovascular Age-Related Macular Degeneration (nAMD). We developed three tools to analyze Optical Coherence Tomography Angiography images: (1) extracting biomarkers such as mCNV area and vessel density using image processing; (2) generating a 3D visualization of the neovascularization for a better view of the affected regions; and (3) applying an ensemble of three white box machine learning algorithms (decision tree, support vector machines and DL-Learner) for nAMD diagnosis. The learned expressions reached 100% accuracy for the training data and 68% accuracy in testing. The main advantage is that all the learned models white-box, which ensures explainability and transparency, allowing clinicians to better understand the decision-making process.

Contrastive Learning in Lesion Detection for Mammography Screening Programs Explained with XAI

Authors: Anca Nicoleta Mărginean

Matrix effect resolution in artificial intelligence-driven bacteria detection

Authors: Anca Nicoleta Mărginean

Accurate detection of bacteria in complex biological, food, and environmental matrices is essential for infection control, antimicrobial resistance (AMR) surveillance, and public health decision-making. However, matrix effects arising from heterogeneous sample compositions remain a major barrier to the reliable deployment of rapid novel detection technologies. These effects can distort analytical signals, reduce reproducibility, and complicate data interpretation, particularly outside controlled laboratory settings. Recent advances in artificial intelligence (AI) offer new opportunities to overcome matrix-induced variability by enabling data-driven signal correction, pattern recognition, and adaptive learning across diverse sample types. In this review, we synthesized emerging strategies from the last 5 years for matrix-effect resolution in AI-assisted bacterial detection systems, including electrochemical, optical, and spectroscopic platforms. We discuss how AI models can integrate biological variability, environmental context, and sensor heterogeneity to improve robustness, analyte discrimination, and even quantitative accuracy, with emphasis on high-risk pathogens, such as those in the ESKAPE group. The workflow for AI-assisted bacterial detection is critically analyzed, beginning with pretreatment of complex samples, processing of raw data to extract the most relevant features, and then continuing with the implementation of the algorithm.

Performance Evaluation of LLMs in Automated RDF Knowledge Graph Generation

Authors: Tudor Cioara, Anghel Ionuț

Cloud systems generate large, heterogeneous log data containing critical infrastructure, application, and security information. Transforming these logs into RDF triples enables their integration into knowledge graphs, improving interpretability, root-cause analysis, and cross-service reasoning beyond what raw logs allow. Large Language Models (LLMs) offer a promising approach to automate RDF knowledge graph generation; however, their effectiveness on complex cloud logs remains largely unexplored. In this paper, we evaluate multiple LLM architectures and prompting strategies for automated RDF extraction using a controlled framework with two pipelines for systematically processing semi-structured log data. The extraction pipeline integrates multiple LLMs to identify relevant entities and relationships, automatically generating subject-predicate-object triples. These outputs are evaluated using a dedicated validation pipeline with both syntactic and semantic metrics to assess accuracy, completeness, and quality. Due to the lack of public ground-truth datasets, we created a reference Log-to-KG dataset from OpenStack logs using manual annotation and ontology-driven methods, enabling objective baseline. Our analysis shows that Few-Shot learning is the most effective strategy, with Llama achieving a 99.35% F1 score and 100% valid RDF output while Qwen, NuExtract, and Gemma also perform well under Few-Shot prompting, with Chain-of-Thought approaches maintaining similar accuracy. One-Shot prompting offers a lighter but effective alternative, while Zero-Shot and advanced strategies such as Tree-of-Thought, Self-Critique, and Generate-Multiple perform substantially worse. These results highlight the importance of contextual examples and prompt design for accurate RDF extraction and reveal model-specific limitations across LLM architectures.

Older people with mild cognitive impairment engaged by social robot-based intervention: Benefits shown in multicountry long term trials (Preprint)

Authors: Tudor Cioara, Anghel Ionuț

BACKGROUND Loneliness and social isolation are some of the several risk factors that contribute to dementia. Technology‑based interventions using socially assistive robots (SARs) and mobile apps may help not only to maintain cognitive functioning but also to support social connectedness and psychosocial wellbeing in older adults with mild cognitive impairment (MCI); however, evidence from multicountry trials remains limited. OBJECTIVE The proof‑of‑concept (PoC) evaluated the benefits of the engAGE platform, a hybrid intervention combining a social robot, mobile app, and wearable activity tracker, designed to support both social connectedness and cognitive functioning in older people with MCI across three European countries. METHODS Older adults with MCI were recruited in Italy, Switzerland, and Norway. The 6‑month intervention combined weekly robot‑guided group sessions with daily tablet use and continuous activity tracker wear at home. Outcomes were assessed at baseline and post-test and included subjective memory complaints as primary outcome (MAC‑Q), and global cognition (MoCA), loneliness (UCLA), quality of life (QoL‑AD and EQ‑5D‑5L VAS), and mental wellbeing (WEMWBS) as secondary outcomes. Also, usability and acceptance were assessed through SUS and UTAUT after 3 and 6 months of intervention. Intra-group and inter‑group differences in change were explored for any dimension to determine the effects of the intervention. RESULTS Of 50 enrolled participants, 44 (36 assigned to the experimental group - EG; and 8 to the control group - CG) completed the final assessment and were included in the analyses. The subjective memory complaints reduced significantly from 26.41 (±2.23) to 25.22 (±3.19) in the EG, whereas remained unchanged in the CG (26.75 ±0.71). MoCA scores remained stable overall (EG: 23.51 ±2.16 to 23.44 ±3.21), with no significant differences between groups. Psychosocial outcomes showed a mixed pattern: in the Italian EG, loneliness decreased significantly (UCLA: 45.18 ± 9.61 to 37.94 ± 6.56), whereas in Switzerland significantly worsened (p=.021). Also, the self‑rated health (EQ‑5D‑5L-VAS) improved significantly in the EG (p=.013), but no significant differences were detected between groups. Overall QoL‑AD and WEMWBS scores remained broadly stable. SUS scores in the EG improved significantly (p=.020) from 58.81 (±18.17) to 68.54 (±18.54), reaching the commonly accepted usability threshold. CONCLUSIONS The engAGE hybrid intervention, combining robot‑guided group activities with home‑based tablet use and activity monitoring, was delivered across three different socio‑healthcare contexts and showed preliminary benefits in subjective memory, selected psychosocial measures, and usability. CLINICALTRIAL ClinicalTrials.gov NCT06302686; https://clinicaltrials.gov/study/NCT06302686 INTERNATIONAL REGISTERED REPORT RR2-10.2196/67601

Asynchronous federated learning with partial weights aggregation for energy consumption forecasting

Authors: Anghel Ionuț, Tudor Cioara

Accurate energy forecasting is essential for grid stability, demand-side management, and efficient renewable integration. However, energy consumption data collected from smart meters may expose sensitive user information, thus raising privacy concerns. Federated Learning (FL) offers a privacy-preserving mechanism for collaborative model training without sharing raw data. However, conventional synchronous FL suffers from training delays caused by heterogeneous client availability and computational capabilities, while frequent exchange of model parameters can lead to communication overheads. To address these challenges, this paper proposes an asynchronous federated learning framework for energy forecasting that enables continuous global model updating without waiting for all clients to complete local training. We introduce a federated asynchronous adaptive aggregation mechanism, where client-specific learning rates are dynamically adjusted based on both update staleness and model performance contribution. A partial aggregation strategy is defined for a Long Short-Term Memory (LSTM) forecasting model that splits the local models' layers, allowing clients to exchange only a subset of the weights with the server. The proposed solution is evaluated using real-world energy consumption data from multiple consumers. Experimental results demonstrate that the proposed asynchronous adaptive strategy outperforms the classic FedAvg approach and maintains prediction accuracy relative to personalised FedAvg, while reducing communication costs. Additionally, the proposed method outperforms the classic FedAsync algorithm across all client groups, with statistically significant improvements in most cases.

Optimizing Appliances Usage for Apartment Buildings Participation in Demand Response Programs

Authors: Anghel Ionuț, Tudor Cioara

This paper presents a bio-inspired optimisation framework that enables apartment buildings to participate more effectively in demand response programmes by coordinating the use of household appliances while preserving residents' comfort. Developed within the context of the DEDALUS project, the proposed approach addresses the complexity of managing multiple apartments with heterogeneous appliances, demonstrating how collective flexibility can support grid stability without requiring the participation of every household. The proposed methodology formulates appliance scheduling as a bi-level optimisation problem, combining a Genetic Algorithm with the Harris Hawks Optimisation (HHO) algorithm to identify both the minimum number of apartments required for demand response participation and the optimal operating schedules for their appliances. The framework incorporates user preferences, appliance flexibility and comfort constraints to generate recommendations that closely match the target consumption profile requested by the distribution system operator while minimising disruptions to residents' daily routines. The approach is validated using data collected from a residential apartment building, demonstrating its ability to closely approximate target energy consumption profiles while involving only a subset of apartments. The experimental results highlight the algorithm's convergence, scalability and effectiveness in balancing energy flexibility with occupant comfort, providing a practical decision-support tool for integrating multi-apartment residential buildings into future smart grid and demand response ecosystems.

Distributed Joint Embedding Predictive Architecture for Energy Forecasting

Authors: Anghel Ionuț, Tudor Cioara

Colonic Polyp Detection with Object Detection Models

Authors: Eugen Richard Ardelean

In recent years, deep learning has been applied more and more to medical image analysis. One such application of deep learning is the automated polyp detection in colonoscopy with the target of reducing miss rates. This study presents a comprehensive evaluation of nine state-of-the-art object detection models for colonic polyp detection: YOLOv8, YOLOv9, YOLOv10, YOLO11, YOLO12, YOLO26, RT-DETR, YOLO-World, and YOLOE. The models were evaluated on three publicly available datasets: CVC-ClinicDB, CVC-ColonDB, and ETIS-LaribPolypDB. All models were trained under standardized conditions using identical hyperparameters and data augmentation strategies to guarantee fair comparison. Performance was evaluated using multiple metrics: mAP@50, mAP@50–95, F1 score, precision, recall, inference time, and computational cost. YOLO11 demonstrated the best overall performance, achieving mAP@50 scores of 0.995, 0.944, and 0.978 on the three datasets respectively, while maintaining the fastest inference time of approximately 150 ms per image and the third-lowest computational cost at 21.3 GFLOPs. Cross-dataset generalization experiments revealed a significant loss of performance, with mAP@50 dropping by 20–40% when models were tested on an unseen dataset, highlighting the challenge of true generalization with limited datasets. Statistical analysis by polyp size showed that while all models achieved F1 scores exceeding 0.95 for large polyps, performance decreased to 0.60–0.85 for small polyps, indicating a limitation in detecting small lesions. The analysis of failure modes showed that missed detections, false positives and boundary errors constitute 60–75% of all failures, suggesting that domain adaptation of object detection models may be required.

Object detection in histology: A multi-dataset benchmark and test-time inference

Authors: Eugen Richard Ardelean

Medical image analysis has become increasingly important for automated medical diagnosis, as well as deep learning. Specifically, object detection models may help in automatically identifying pathological structures and features. This study presents a comprehensive comparative analysis for object detection tasks in histological images of the latest models including the YOLO (You Only Look Once) architectures, from YOLOv8 to the recently introduced YOLOv12. These models were evaluated alongside alternative architectures including RT-DETR, YOLO-World, and YOLOE across five diverse histology datasets: BCNB, Nuclei, TNBC, MoNuSAC, and CryoNuSeg. The experimental analysis employed standardized training protocols with consistent hyperparameters and data augmentation strategies, evaluating the performance through multiple metrics, inference time, and computational cost. The results obtained on the five datasets indicate that YOLOv11 consistently showed a strong performance across multiple datasets, however the newly introduced attention mechanisms of YOLOv12 show good performance, despite the model having slightly lower overall performance. Specialized variants like YOLOE demonstrated promising results for specific applications, while RT-DETR showed poor performance on smaller objects, which are typical in histological images. Statistical analyses indicate that YOLOv11 indeed has the best performance but that all models have a poor performance on objects of small sizes; moreover, the most common cases of failure are background false positives and missed detections. This comprehensive evaluation provides insights for the current state of object detection architectures for clinical histopathology applications and establishes benchmarks for future avenues of research in automated medical image analysis. In addition to the multi-model benchmark, we propose Test-time Graph Similarity Propagation (TGSP), a test-time self-supervised refinement that uses ResNet50 deep features to build a k-NN similarity graph over detections and performs label propagation to re-score predicted boxes. TGSP replaces TSBP's iterative Earth-Mover matching with adaptive per-class quantile thresholds and graph-based label propagation, eliminating K-means hyperparameters and better scalability. Our analysis on histology datasets TGSP consistently matches or improves F1 relative to both a fixed 0.5 threshold and TSBP, with the biggest gains when base-model confidence calibration is poor.

Multi-Agent Coverage Using Multiplicatively Weighted Energy Voronoi Partitions

Authors: Lucian Bușoniu

We propose a coverage control method for energy-constrained multi-agent systems with single-integrator agent motion, in which agent energy discharges and recharges at constant rates. Differently existing methods, a Multiplicatively Weighted Energy Voronoi (MWEV) partition ensures that each agent’s coverage region varies with energy in such a way that it vanishes when energy drops to a reserve level sufficient to reach a charging station. We show that the optima of an energy-aware coverage objective are the dynamically evolving MWEV centroids. A generalized Lloyd algorithm provably drives agents with remaining battery to these centroids, and reserve-level agents to their charging stations, under a switched two-timescale model with fast motion and slow energy dynamics. Agents repeatedly pause coverage as they get depleted, and then resume coverage upon recharging. This happens arbitrarily many times, leading to an infinite-horizon coverage setting. The method works well in simulations, where we also apply an alternative technique with a different problem formulation. We contrast coverage performance and average agent downtime between the two methods; e.g., downtime is 17.70% of the experiment duration for our MWEV technique, compared to 28.5% for the alternative. In addition, the robustness of the proposed approach is investigated under nonlinear battery discharge dynamics, position errors, and delayed energy information.

Fast Neural-Network Approximation of Active Target Search Under Uncertainty

Authors: Lucian Bușoniu

We address the problem of searching for an unknown number of stationary targets at unknown positions with a mobile agent. A probability hypothesis density filter is used to estimate the expected number of targets under measurement uncertainty. Existing planners, such as Active Search (AS) and its Intermittent variant (ASI), achieve accurate detection but require costly online optimization. To reduce online computation, we propose to use a convolutional neural network to approximate AS or ASI decisions through direct inference. The network is trained on AS/ASI data using a multi-channel grid that encodes target beliefs, the agent position, visitation history, and boundary information. Simulations with uniform and clustered target distributions show that the network achieves detection rates comparable to AS or ASI while reducing computation by orders of magnitude.

Value iteration with stopping criterion: finite iterations, stability, and near-optimality guarantees

Authors: Lucian Bușoniu

Value iteration (VI) is a cornerstone of dynamic programming that allows computing near-optimal feedback laws for general plant dynamics and cost functions. In practice, however, it must be stopped after finitely many iterations. This raises the question of when to stop the algorithm so that the resulting policies and value functions achieve desirable properties, like given near-optimality bounds and stability. In this context, we study deterministic, discrete-time systems with infinite-horizon (possibly discounted) costs whose inputs are generated by VI. We equip VI with a generalized stopping criterion that encompasses existing choices while allowing new ones. Our aim is to analyze the properties of the policies and value functions at the final iteration. Under mild assumptions, we first show that VI indeed terminates in a finite number of iterations. We then establish that the final policies are stabilizing by properly designing the stopping criterion, and derive explicit near-optimality bounds characterized by this choice. These results offer a design framework for the stopping criteria that balances computational effort with stability and performance guarantees.

Segmentation of the Retinal Vascular Network and Biomarker Quantification in OCTA Imaging

Authors: Mihnea Jurca

Forest Inspection Dataset: A Synthetic UAV Dataset for Semantic Segmentation of Forest Environments

Authors: Sergiu Nedevschi

This work describes the Forest Inspection dataset, a synthetic aerial image collection designed for semantic segmentation of forest environments with an emphasis on UAV-based forest inspection. The dataset consists of high-resolution RGB images paired with dense pixel-level semantic labels covering 11 classes, including deciduous trees, coniferous trees, fallen trees, ground vegetation, dirt ground, rocks, sky, buildings, fences, and vehicles. Images were generated in AirSim using a photorealistic virtual forest environment and captured with simulated UAV flights at three altitudes (30 m, 50 m, 80 m) and three camera pitch angles (0°, 60°, 90°) to reproduce diverse observation conditions, under two weather settings: sunny and overcast. Each data sample includes the corresponding UAV pose metadata for spatial context. The dataset is provided in standard image and annotation formats, accompanied by a description of the scene configuration and acquisition parameters. This resource is intended to support the development and evaluation of semantic segmentation models and other computer vision methods for UAV-based forest scene understanding and inspection.

Improving Counting Accuracy of Postdisaster Visual Question Answering for Remote Sensing

Authors: Sergiu Nedevschi

In post-disaster damage assessment, visual question answering (VQA) systems are essential in identifying the severity and scope of damage. However, counting-related tasks, such as determining the number of vehicles and flooded buildings, remain a significant challenge for current deep learning models. To address this issue, we propose DeVANet (DeBERTa Vision Attention Network), a novel architecture aimed at enhancing counting accuracy in VQA for post-disaster scenarios. We leverage DeBERTa for language modeling and introduce an innovative image embedding module, where local-global attention guides Vision Mamba features to achieve precise extraction of both small and large objects. Our fusion mechanism employs self-attention for both text and image data, followed by bidirectional cross-attention and co-attention to enhance multimodal integration. We tackle VQA as both a classification and regression problem by employing separate MLPs for each task: one handling discrete class predictions and the other generating continuous values for counting tasks. A joint loss function, combining weighted cross-entropy and negative binomial loss, ensures optimized performance across both tasks. Extensive experiments on the FloodNet and RescueNet datasets demonstrate that DeVANet achieves significant improvements in counting accuracy and overall VQA performance compared to state-of-the-art works, supported by detailed ablation studies that validate the effectiveness of each component in the architecture.

Memetic-based Coordination of Distributed Storage Units Flexibility for Congestion Management

Authors: Tudor Cioara, Anghel Ionuț

P2P Energy Trading Coordination in Interconnected Microgrid Systems

Authors: Tudor Cioara

Edge-Oriented Orchestration of Energy Services Using Graph-Driven Swarm Intelligence

Authors: Tudor Cioara, Vasile Ofrim

As smart grids increasingly depend on IoT devices and distributed energy management, they require decentralized, low latency orchestration of energy services. We address this with a unified framework for edge fog cloud infrastructures tailored to smart energy systems. It features a graph based data model that captures infrastructure and workload, enabling efficient topology exploration and task placement. Leveraging this model, a swarm-based heuristic algorithm handles task offloading in a resource-aware, latency sensitive manner. Our framework ensures data interoperability via energy data space compliance and guarantees traceability using blockchain based workload notarization. We validate our approach with a real-world KubeEdge deployment, demonstrating zero downtime service migration under dynamic workloads while maintaining service continuity.

Performance evaluation of large language models for automated knowledge graph generation

Authors: Tudor Cioara, Anghel Ionuț

Cloud systems generate large and heterogeneous raw data logs containing information related to infrastructure, application, and security. Transforming these logs into RDF triples enables their integration into knowledge graphs, improving interpretability, root-cause analysis, and cross-service reasoning. Large Language Models (LLMs) offer a promising approach to automate RDF knowledge graph generation. However, their effectiveness on complex cloud logs remains largely unexplored. In this paper, we evaluate multiple LLM architectures and prompting strategies for automated RDF extraction using a controlled framework with two pipelines for systematically processing semi-structured log data. The extraction pipeline integrates multiple LLMs to identify relevant entities and relationships, automatically generating subject-predicate-object triples. These outputs are evaluated using a dedicated validation pipeline with both syntactic and semantic metrics to assess accuracy, completeness, and quality. Due to the lack of public ground-truth datasets, we created a reference Log-to-KG dataset from OpenStack logs using manual annotation and ontology-driven methods, enabling objective baseline. Few-Shot learning achieved the best overall performance in our evaluation, with LLaMA achieving a 99.35% F1 score and 100% valid RDF output while Qwen, NuExtract, and Gemma also perform well under Few-Shot prompting, with Chain-of-Thought approaches maintaining similar accuracy. One-Shot prompting offers a lighter but effective alternative, while Zero-Shot and strategies such as Single-Prompt Tree-of-Thought, and Generate-Multiple-Then-Vote showed consistently lower performance. The top-performing LLMs, LLaMA and Qwen, with One-Shot and Few-Shot prompting were evaluated on a different dataset, to assess transferability, performance stability and efficiency in terms of latency, token generation, and GPU memory usage. Our work highlights the importance of contextual examples and prompt design for accurate RDF extraction and reveal model-specific limitations across LLM architectures offering practical guidelines and a foundation for future knowledge graphs extraction from unstructured data research.

Energy forecasting under missing data: Comparative evaluation of augmented representations and decoder-only time-series imputation

Authors: Tudor Cioara, Mircea Gabriel Antonesi, Anghel Ionuț

Data-related issues, including missing values and irregular measurements, challenge the accuracy of short-term energy forecasting in smart grids. In data-scarce scenarios, two approaches are commonly considered, but their strengths and weaknesses are not fully mapped. Embedding-based models learn joint representations from heterogeneous data, compensating for the lack of time-series measurements via additional contextual or external sources, whereas imputation pipelines restore temporal continuity but may smooth variability or produce implausible values. To address these limitations, we propose a unified forecasting framework for energy systems that integrates a shared Temporal Fusion Transformer prediction with a controlled degradation protocol to simulate realistic missing-data patterns. This enables a fair and systematic comparison between two pipelines: a representation-augmented learning and decoder-only time series imputation. The former integrates TS2Vec temporal embeddings and BERT-based static contextual representations to provide a richer forecasting space without explicit reconstruction of missing values. The latter uses a Chronos-2 model to reconstruct missing time-series segments, followed by physics-based correction to enforce physically plausible outputs. We evaluate both pipelines under a controlled data degradation protocol to map the trade-offs between representation learning and data continuity restoration through imputation. We use real-world non-residential building electricity consumption and wind generation datasets. The imputation-based pipeline achieves a mean sMAPE of 10.14% and MAE of 8.43 kWh across 100 buildings, compared to 12.11% and 10.89 kWh for the representation-based approach (p<0.01, p<0.01, p<0.01). On the wind generation dataset, imputation also improves predictive accuracy (R²=0.870 vs. R²=0.794). However, representation-based models remain competitive in scenarios with irregular, spike-dominated, or event-driven consumption patterns where imputation provides limited additional benefits.