Abstract
Peer-reviewed paper 7-108-26 The energy transition creates significant challenges for long-term planning in the buildings sector, which accounts for around 35% of EU greenhouse gas emissions. Frameworks such as the Energy Performance of Buildings Directive promote electrification and the deployment of heat pumps, yet the representation of domestic hot water (DHW) technologies in long-term energy models remains challenging due to limited data. This paper examines the integration of DHW heat pumps into the OSeMOSYS-based national energy system model, using Portugal as a case study. Key modelling gaps are identified in the DHW demand, technology performance, and economic parameters. National statistics provide aggregated residential energy data, while detailed end-use information and hourly DHW profiles remain scarce. To address these limitations, the modelling framework reconstructs DHW demand from national energy balances, derives stylised demand profiles from literature proxies, and represents heat pump performance using seasonal coefficients of performance (COP). Technology costs, including capital investment, maintenance, and electricity prices, are parameterised using market data and European literature benchmarks. Instead of exploring numerous scenarios, this study quantifies how key modelling assumptions, particularly cost parameters and COP variations, affect heat pump deployment, system costs, and electricity demand. Results show that cost assumptions influence heat pump deployment, while variations in COP affect electricity demand, with implications for grid planning. The outcomes demonstrate that modelling assumptions can lead to substantial variations in system outcomes, even under consistent decarbonisation constraints. The study emphasises the importance of transparency when modelling under data scarcity and illustrates a pragmatic approach to representing DHW heat pumps in national energy system planning models.

Abstract
Accurate time series forecasting is crucial across various domains, yet traditional models that rely solely on numerical data often struggle to capture complex patterns in dynamic environments. This work proposes a novel multimodal forecasting framework that integrates visual and numerical data to enhance predictive performance. The framework leverages a FT-Transformer for temporal data processing and a TIMM-based convolutional network for visual data extraction. A hybrid fusion strategy combines these modalities, enabling the model to capture complementary information that improves forecasting accuracy. Empirical evaluations on the M4 dataset demonstrate that the multimodal model consistently outperforms unimodal approaches, achieving up to a 7.0% reduction in Normalized Root Mean Squared Error across multiple forecast horizons. The proposed framework also incorporates an automated training pipeline powered by Optuna, ensuring efficient hyperparameter tuning and scalability across diverse datasets. These results highlight the effectiveness of multimodal integration in advancing time series forecasting performance. © 2026 IEEE.

Abstract
This paper explores advanced methods to enhance medical image captioning for X-ray images, leveraging clinically significant labels to generate precise, informative descriptions. Building on state-of-the-art baseline models, this study systematically evaluates various image encoders and text decoders to optimise the clinical relevance and linguistic quality of generated captions. To ensure rigorous and meaningful evaluation, domain-specific metrics tailored for medical imaging are employed. Furthermore, the research investigates efficient fusion strategies that incorporate extracted clinical labels into the captioning pipeline without sacrificing model performance. To this end, we introduce X-CapFusion, a lightweight, label-conditioned, multimodal alignment framework for medical image captioning. X-CapFusion integrates label-aware fusion mechanisms with scalable encoder-decoder architectures, enabling the generation of semantically rich and clinically grounded descriptions. The framework is evaluated on multiple datasets, showing superior caption accuracy, informativeness, and robustness in low-resource medical imaging scenarios. By leveraging structured label guidance and modality alignment, it reduces computational costs while preserving factual consistency and semantic depth, achieving competitive performance on both clinical and general language metrics and remaining suitable for real-world deployment. With a best RadCliQ-v1 score of 1.251 on MIMIC-CXR, X-CapFusion demonstrates its efficacy in generating clinically accurate, semantically coherent medical captions while maintaining high computational efficiency. Overall, it achieves competitive performance across both traditional and domain-specific metrics. Source code of the project available at https://github.com/magdamorim/X-CapFusion.

Abstract

Abstract
Accurate infrastructure-based UAV localization remains challenging in the presence of occlusions, clutter, and limited observability from single sensing modalities. We present a distributed multi-station tracking framework that fuses mmWave radar and monocular vision to achieve robust 3D position and velocity estimation. Building upon a prior single-station radar-vision tracker, we extend the approach to a network of three portable observation stations, each performing local multi-hypothesis tracking with uncertainty-aware Kalman filtering. Vision measurements provide angular constraints that improve radar data association and mitigate clutter-induced artifacts. Instead of transmitting raw detections, each station communicates a compact state estimate and covariance to a central fusion node, where an information-form filter produces a globally consistent estimate. The system is validated in indoor flight experiments using motion capture ground truth, while remaining fully independent of it during estimation. The fused solution achieves a 3D RMSE of 0.2342 m and improves robustness against degraded individual station estimates. These results highlight the potential of distributed radar-vision sensor networks for scalable and reliable infrastructure-based UAV localization.

Abstract
Maintaining volunteer engagement without sacrificing annotation quality is a central challenge for citizen-science and crowdsourcing systems. This paper presents a one-week quasi-experimental field trial of JellyfishGO! that tests how competitive and cooperative Games with a purpose (GWAP) mechanics affect motivation and data outcomes, involving 64 players (n = 64) organized in 32 two-person teams. We address two research questions (RQs): how motivation varies across Bartle player types (RQ1), and which mechanics best optimize engagement and data quality (RQ2). Results show that a Hybrid configuration (team matches + light inter-team leaderboard) substantially increased validated capture throughput versus solo-competitive play (Incidence Rate Ratio (IRR) ˜ 1.79, 95% Confidence Interval (CI) [1.37, 2.33]) and that this uplift was broadly distributed across Bartle player types. © 2026 IEEE.