Decarbonizing the heating sector is central to achieving the energy transition, as heating systems provide essential space heating and hot water in residential and industrial environments. A major challenge lies in effectively pro filing large clusters of buildings to improve demand estimation and enable efficient Demand Response schemes. This paper addresses this challenge by introducing a novel unsupervised machine learning framework for clus tering residential heating load profiles, focusing on natural gas space heating and hot water preparation boilers, while analyzing five different dimensions: boiler usage, heating demand, weather conditions, building charac teristics, and user behavior. Three distance metrics, Euclidean Distance, Dynamic Time Warping, and Derivative Dynamic Time Warping, are applied and evaluated using established clustering indices. The proposed method is assessed considering 29 residential buildings in Greece equipped with smart heating controllers throughout a calendar heating season (i.e. 210 days). This study demonstrates that Dynamic Time Warping is demonstrably the most suitable metric. A subsequent correlation analysis of the clustering results from each dimension reveals strong, time-dependent relationships between boiler usage, heat demand and temperature, identifying them as the most important and correlated directions. These findings shed light on heating load behavior, establishing a solid foundation for developing more targeted and effective demand response programs.

The decarbonization of the building sector remains a critical component of climate action, as buildings are responsible for a significant share of global CO2 emissions and energy consumption. Despite numerous efforts, existing approaches often fall short in addressing the complexity of building systems, user-specific characteristics, budgetary constraints and real-world decision-making processes. This study introduces a Mixed-Integer Linear Programming (MILP) methodology embedded within a Decarbonization-as-a-Service (DaaS) framework that supports stakeholders in identifying optimal retrofit strategies tailored to each building’s profile. The framework integrates detailed user-provided inputs—covering passive and active systems, energy usage patterns and renovation budgets, with a structured optimization model that selects among predefined decarbonization actions. These include energy-efficient envelope upgrades, replacements for HVAC and electrical appliances, and installations of photovoltaic (PV) and battery storage systems. The model aims to minimize annual CO2 emissions while adhering to cost and technical feasibility constraints. The approach is validated through two case studies: one of a single-family residential prototype building model and one of a multi-residential building complex located in Vienna, Austria. Results are compared against the high-fidelity simulators EnergyPlus and SIM-VICUS confirming the accuracy of the proposed energy assessment models. Representative renovation scenarios are then evaluated across different budget levels, showcasing the framework’s ability to achieve up to near-complete decarbonization (97.44% for the Austrian case). These results demonstrate the methodology’s ability to deliver effective, scalable decarbonization pathways aligned with EU climate targets and renovation wave objectives.

Hierarchical forecasting plays a critical role in advancing building energy systems by ensuring coherence across different levels of aggregation, such as individual rooms, building floors, and entire facilities. While traditional reconciliation approaches, such as bottom-up and top-down methods, offer simplicity, they often ignore valuable information across hierarchical levels, leading to myopic and biased forecasts. Recent advances have introduced machine learning into hierarchical forecasting to improve accuracy and allow for more flexibility. In this study, we propose a novel reconciliation framework that builds on neural networks and is tailored to electricity demand forecasting in buildings. We introduce a set of loss functions that penalize incoherence across various hierarchical relationships, including discrepancies between adjacent levels and between the top and bottom levels. Our method employs a global model that simultaneously generates and reconciles forecasts across the hierarchy, thereby eliminating the need for independent base forecasts and enabling cross-series learning. We validate our method using electricity consumption data from a modern office building in Helsinki, Finland. Results demonstrate that our approach outperforms standard reconciliation methods and achieves a significant loss reduction in terms of both accuracy and coherence in complex, multi-level building settings. The results translate into practical benefits that can be used to improve building energy management by enabling more accurate load shifting, enhanced scheduling of HVAC operations and improved utilization of on-site renewable generation, all of which highlight the utility of the approach in complex, real-world energy forecasting applications.

Volatility in the modern world and electricity Day-Ahead Markets (DAMs) usually makes long-term historical data irrelevant or even detrimental for accurate forecasting. This study directly addresses this challenge by proposing a novel forecasting paradigm centered on extremely short training windows, ranging from 7 to 90 days, to maximize responsiveness to recent market dynamics. This volatility-driven approach intentionally creates a data-scarce environment where the suitability of deep learning models is limited. Building on the hypothesis that shallow machine learning models, and more specifically boosting trees, are better adapted to this reality, we evaluate four models, namely LSTM with feed-forward error correction, XGBoost, LightGBM, and CatBoost, across three European energy markets (Greece, Belgium, Ireland) using feature sets derived from ENTSO-E forecast data. Results consistently demonstrate that LightGBM provides superior forecasting accuracy and robustness, particularly when trained on 45–60 day windows, which strike an optimal balance between temporal relevance and learning depth. Furthermore, a stronger capability in detecting seasonal effects and peak price events is exhibited. These findings validate that a short-window training strategy, combined with computationally efficient shallow models, is a highly effective and practical approach for navigating the volatility and data constraints of modern DAM forecasting.

Traditional energy forecasting solutions rely on task-specific supervision and energy asset representations, limiting transferability and the ability to capture general temporal dynamics across heterogeneous assets. We address this by proposing a distributed Joint Embedding Predictive Architecture (JEPA) for self-supervised learning from heterogeneous energy time-series. The framework predicts latent representations of masked temporal segments while integrating temporal observations and contextual information within a shared embedding space. To prevent representation collapse, training combines a latent-space predictive objective with covariance and temporal variance regularization. The evaluation was conducted on energy consumption and generation datasets under data-degradation scenarios and compared with a Transformer forecasting baseline. The learned representations remained stable (cosine similarity ≈0.98; effective rank 185–235). JEPA achieved performance comparable to a Transformer on building energy data, higher R² in 3/5 consumer clusters, and outperformed the baseline on 9/10 unseen PVs (R²=0.73–0.88 vs. <0.45), while showing greater robustness to missing data.

Real-time residential energy monitoring at scale demands pipelines capable of handling the irregularities endemic to live smart meter feeds. This paper presents the architecture and initial deployment results of a continuous energy monitoring platform integrated directly with the API of Heron, a major Greek electricity supplier, ingesting smart meter data from 141 residential homes of the Heron Living Lab. An automated FastAPI-based ingestion pipeline feeds a time-series database and React-based monitoring dashboard, enabling analytics that would be precluded by offline, batch-oriented workflows. A dedicated data quality module enforces measurement reliability through missing value detection, sampling fidelity validation, and a two-stage anomaly detection scheme: a per-home rolling zscore detector provides real-time flagging at each polling cycle, while a per-home autoencoder, trained exclusively on clean windows and thresholded at the 95th percentile of reconstruction error, captures temporally structured deviations unsuitable for statistical baselines. Characterization of the live stream over the full deployment period (August 2020 - present) reveals a fleetwide median within-window completeness of 78.3%, pronounced inter-household heterogeneity in both missingness and gap structure, and a peak operational cohort of 106 concurrently active homes. These findings underscore data quality enforcement as a structural pipeline requirement rather than a preprocessing convenience. The validated stream feeds a downstream analytics layer for flexibility estimation and appliance-level disaggregation.

The accelerated penetration of distributed renewable energy sources in energy communities has shifted the operational bottleneck from generationto coordinated demand-side management. This paper presents FLEXCOMM, a web platform supporting self-consumption optimization in energy communities through two complementary services. Its public, community-level service aggregates community demand and PV park production into 24-hour measured and forecast curves, computes self-consumption indicators, and translates surplus windows into natural-language load-shifting suggestions. Its authenticated, member-level service delivers comfort-aware HVAC scheduling that operates identically under cooling- and heating-dominated conditions, coupling unsupervised HVAC disaggregation of the aggregated smart-meter signal with regime-aware indoor temperature and humidity forecasting and a mixed-integer linear programming scheduler with a machine-learning-based comfort-repair loop under thermal-comfort constraints. Both services share a containerized FastAPI/React architecture with role-based access; each layer is independently deployable, connects to existing metering, sensing, and weather infrastructure through documented REST APIs, and serves forecasting models as pluggable components, so new communities can be onboarded without modifying the codebase. FLEXCOMM operates in production for the Chalki energy community (Greece), with the member-level service replicated for the BCS Energia energy community (Hungary), demonstrating community-wide and household-level self-consumption analytics from a single, non-intrusive platform without appliance-level sub-metering.

The integration of renewable energy and dynamic pricing in European electricity systems requires innovative approaches to manage residential flexibility. This paper presents an elasticity-aware clustering and optimisation framework to jointly model behavioural and asset-based flexibility, addressing gaps in current demand-side management strategies. The methodology comprises three tools: (i) causal inference to estimate consumers’ price elasticities, isolating the impact of price variations on consumption; (ii) unsupervised clustering using both elasticity estimates and electric water heater (EWH) load profiles, capturing heterogeneity in flexibility; and (iii) cluster-level optimization for EWH scheduling, minimizing energy usage while preserving user comfort. Simulation results demonstrate measurable reductions in electricity usage across clusters (around 75%). The framework supports aggregators and retailers in designing differentiated market strategies, enhancing coordination in future flexibility markets. By bridging behavioural and operational flexibility, this approach offers relevant insights for residential demand-side management.

The increasing penetration of distributed energy resources (DERs) in distribution networks is turning traditional consumers into prosumers through active participation in energy communities (EnCs). While this shift supports decentralization, it also creates new challenges-especially in coordinating electricity markets across transmission and distribution levels. This study proposes a coordinated transmission-distribution market clearing model that maximizes EnC profitability through the integration of distributed photovoltaic (PV) systems, while ensuring system reliability. The model includes line capacity constraints and simulates both congestion-prone and congestion-free scenarios across all seasons. We model the EnC as a virtual power plant (VPP) operating under a virtual net billing (VNB) scheme, and analyze how varying the number and location of $\mathbf{P V}$ installations affects economic performance. Profit gains are assessed by combining VNB revenues with changes in locational marginal prices caused by PV injections. Results show that in congestionprone conditions, higher PV deployment can reduce revenues due to export limitations and sustained low prices. However, rather than discouraging investment, these price signals help guide efficient system evolution-limiting oversupply in saturated areas and attracting new demand or flexible loads to absorb surplus energy. In congestion-free scenarios, the EnC can export nearly all generated energy year-round, with profits scaling directly with PV capacity and driven largely by VNB revenues. These findings highlight the importance of coordinated planning and adaptive market design in capturing the full value of distributed PV.

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 8.43kWh across 100 buildings, compared to 12.11% and 10.89kWh for the representation-based approach (p<0.01 p<0.01 p<0.01). On the wind generation 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.

The application of Artificial Intelligence (AI) in the energy sector offers new opportunities for developing flexible, efficient, and sustainable infrastructures. Nevertheless, real-world deployment is still constrained by the lack of large-scale, integrated environments that can evaluate advanced algorithms under realistic operating conditions while ensuring regulatory compliance. This paper presents EnerTEF (which stands for Energy Testing and Experimentation Facility), a federated platform for testing and experimentation in the energy sector designed to address this gap. We introduce a unified TEF architecture that enables full-stack evaluation of intelligent systems, including predictive modeling, optimization, learning under data distribution shifts and federated learning across geographically distributed sites. The framework integrates high-fidelity digital twins, a privacy-preserving data exchange framework and regulatory sandboxing to support transparent, explainable and robust AI development. EnerTEF demonstrates how such a framework can be deployed in critical energy domains through three real-world scenarios including short-term hydropower generation forecasting, coordination between distribution network operators and distributed energy resources and real-time optimization of self-consumption for municipal buildings. Results show that EnerTEF effectively enables the development of novel AI models, improves cross-context generalizability and supports innovation for complex energy infrastructures, ultimately creating a practical, scalable path for addressing different energy-related problems and heterogeneous data.

Residential heating can provide demand-side flexibility by temporarily reducing heat demand while maintaining acceptable indoor conditions. However, apartment-level flexibility is difficult to quantify because it depends on heating demand, thermal response, weather conditions, thermostat settings, and comfort constraints. This paper presents a hybrid data-driven and physics-informed framework for assessing downward heating flexibility in district-heated residential buildings. The framework combines short-term heating demand forecasting, apartment-specific thermal response modelling, and comfort-constrained simulation. Candidate heating reductions are evaluated using indoor temperature, PMV–PPD, relative humidity, and CO2 constraints, and the maximum admissible reduction is expressed through a Heating Flexibility Indicator (HFI) and the corresponding energy flexibility in kWh. The framework is evaluated using hourly data from 17 apartments in two Danish residential social housing blocks. Ridge regression, Piecewise Ridge regression, and a physics-informed Mamba model are compared for baseline forecasting, with the Mamba model achieving the strongest overall performance and R2 values above 0.8 for most apartments. For the case-study day, the framework identifies 27.34 kWh of admissible heating reduction across the portfolio, with 18.12 kWh from Block 1 and 9.22 kWh from Block 4. The seasonal analysis shows that spring provides the highest average HFI, whereas autumn provides the largest energy flexibility. The results indicate that residential heating flexibility varies across apartments, blocks, and seasonal operating conditions, supporting selective rather than uniform DR activation.

While transformer models have achieved strong performance in time series forecasting, they often underestimate peak demand and generation events, producing over-smoothed predictions. Existing diffusion-based forecasting approaches face similar challenges, as they mostly rely on Gaussian noise assumptions and lack explicit spike-aware conditioning, which limits their ability to model extreme events and heavy-tailed energy time series. To address these limitations, in this paper, we propose a hybrid diffusion transformer model that explicitly integrates spike-aware information into the generative process for energy forecasting. Unlike existing approaches that model uncertainty and event detection separately, our method couples with the diffusion transformer, a Mamba-based energy time series spike detector and uses an adaptive Johnson-SB distribution to better capture the skewed and heavy-tailed characteristics of energy data. On top, we introduce three novel integration strategies that progressively incorporate spike information into the forecasting pipeline, ranging from post-inference correction to direct conditioning of the diffusion dynamics and stochastic noise generation. Experimental results on electricity demand, heat demand, and wind generation data prove that the proposed model improves forecasting performance. The post-processing adjustment strategy reduced spike-point RMSE by 11.1% and 16.4% on the electricity and heat demand data, while latent modulation achieved the best overall performance on wind generation forecasting with a R2 score of 0.912 and a RMSE improvement of 47.12% over the vanilla transformer baseline.

The ongoing evolution of energy systems is shifting conventional power grids toward smart grids, where advanced metering infrastructure and smart meters provide detailed building-level data. However, many buildings (residential or not) are characterized by historical load data scarcity, which constrains the accuracy of short-term load forecasting models, necessary for participating in energy management schemes. Transfer Learning (TL) provides a compelling approach by enabling pre-trained models to be adapted to data scarce environments, effectively leveraging knowledge acquired from data-rich domains. This study explores a systematic experimental framework consisting of four established baseline models alongside the state-of-the-art Temporal Fusion Transformer. Each model is pretrained on a ”source” building (data-rich) and subsequently transferred to target (data-poor) buildings of varying domain similarity, simulating scenarios of data scarcity. To enable effective knowledge transfer, three to five parameter-based transfer learning techniques were tested per baseline model, including full fine-tuning and partial layer adaptation, while a total of three novel fine-grained knowledge transfer methods were applied on the Temporal Fusion Transformer targeting its variable selection, recurrent and attention components. Experimental results on a real-world dataset containing student dormitory buildings from three different geographical locations, demonstrate that TL consistently enhances forecasting accuracy across all target domains, with the largest benefits observed under severe data scarcity, where the best performing TL configurations achieved an RMSE reduced by 10-20% compared to base models. Even in settings with larger domain gaps or longer historical records, TL still yielded stable improvements in all metrics across all models while the Temporal Fusion Transformer notably yielded around 20% reduced RMSE and 41% increased R2 Score. Moreover, it surpassed the strongest TL baselines, delivering up to 22% additional improvement in R2 Score and consistent reductions in RMSE and MAE across domains. These findings highlight both the effectiveness of TL in mitigating data scarcity and the robustness of Temporal Fusion Transformer in extracting temporal and contextual patterns, confirming its value for cross-domain and limited-data load forecasting.

Single-task fine-tuning of graph neural networks (GNNs) for power grid problems exhibits a systematic failure mode: models that achieve the lowest in-distribution error degrade the most under topology shift. We term this topology overfitting: the tendency of task-specific gradient signals to encode relational structure particular to the training topologies rather than the underlying physics, causing models to fail on unseen grids despite strong in-distribution performance. To expose and address this failure mode, we introduce MxGPS (Multiplex GPS), a multiplex graph transformer that runs K task-specialised GPS branches over a shared node encoder, jointly trained on Static State Estimation (SSE) and AC Power Flow (PF) via a self-supervised pre-training and multi-task fine-tuning protocol, with a cross-branch attention module evaluated in ablation. The joint SSE+PF objective forces the shared encoder to simultaneously satisfy complementary gradient signals, preventing it from overfitting to topology-specific relational structure. Under a 3-fold sliding-window cross-validation spanning four unseen topologies (14-, 24-, 162-, and 300-bus), MxGPS attains 0

The building stock in the European Union (EU) is in urgent need of energy efficiency renovations. Initiatives such as the EU Renovation Wave Strategy, the revised Energy Performance of Buildings Directive (EPBD) and the New European Bauhaus aim to promote deep renovations and unlock energy efficiency investments, however a wider outreach is still needed. The proposed study presents the LiveBetter project which aims to accelerate the adoption of deep renovations across Europe by establishing the Standardized Common European Digital Buildings Hub, which integrates a high-performance, Data Space-compliant data-sharing mechanism for quality assurance in the building sector. The LiveBetter Hub also provides an arsenal of 15 market-ready, Artificial Intelligence (AI) based services destined for accurate, reliable, dynamic Energy Performance Certificates (EPCs), digital building inspections, digitaltwin-based performance-versus-comfort building management, renovation planning, financing, and flexible, demand-responseready buildings, while remaining compliant with the EU Building Stock Observatory (EU BSO). These solutions will be validated through six unique Business Cases, which will be tested in six End-user Groups covering key stakeholders across the building value chain such as building owners, construction companies, financial institutions, occupants, energy utilities, local authorities and academia. Beyond technical innovation, the project further ensures market uptake through innovative financing schemes, policy support, business models and capacity-building actions, with particular attention to non-technical stakeholders through a Community of Practice and Building Clinics.

Household energy demand data are essential for designing load-shifting strategies, storage solutions and demand response programs. This is the first publicly available dataset that integrates multiple household-level energy and mobility metrics. These include grid imports and exports (30-minute intervals), rooftop photovoltaic (PV) production (30-minute intervals), EV charging sessions, and detailed journey logs (start time, end time, distance, and duration) from the same households of a pilot energy community located on the Dingle Peninsula, Ireland. The dataset includes four volunteer households, each equipped with a 2.1 kWp rooftop PV system, 5 kWh Sonnen battery, a Hyundai Kona Electric (64 kWh), a Pulsar Plus EV charger (7.4 kW AC), and a Mitsubishi Electric Ecodan air-source heat pump.While grid and PV data span three years, EV mobility logs cover February 2021 to January 2022. Battery telemetry and heat pump demand data are available for the final six to eight months of the observation period (mid-2021 to early 2022), enabling detailed whole-home analysis for that specific window.

This paper introduces Energy-VERITIES, a framework for modeling epistemic risk and trust towards the integration of generative AI and large language models (LLMs) within the energy digital spine ecosystem. It defines epistemic risk zones using a set-theoretic Venn model over AI outputs ($\boldsymbol{A}$), expert ($E$) and non-expert ($N$) beliefs, internet sources ($D \_I$), and verified domain facts ($D$), with a critical risk zone formalized as (D_I $\cap A \cap E \cap N$) - D, representing hallucinated outputs that appear credible due to overlapping AI, expert, and stakeholder beliefs but diverge from verified domain facts. To assess energy stakeholderAI interaction, we propose a set of Trust Performance Indicators (TPIs) that quantify epistemic outcomes from the user's perspective, extending the confusion matrix and the previously devised VIRTSI model, which characterizes dynamic human trust states through a finite state automaton informed by user validation behavior. These include, for example, MJO-HRAR (Misconceptually Justified Overtrust), where users accept false outputs after flawed validation, and BOV-HRAR (Blind Overtrust), where no validation occurs. An empirical study of ChatGPT's energy advice in Greece illustrates the framework's diagnostic value, revealing that 39 % of validated responses were still false, over 30 % of hallucinations were accepted without validation, and correct answers were often dismissed. These findings indicate the need for improvements in interface-level explainability, validation support, and targeted AI literacy for energy stakeholders. The Energy-VERITIES framework offers a foundation for identifying epistemic risks and guiding the trustworthy integration of generative AI and LLMs in high-stakes domains, like energy.

Federated Learning is transforming electrical load forecasting by enabling Artificial Intelligence (AI) models to be trained directly on household edge devices. However, the prediction accuracy of federated learning models tends to diminish when dealing with non-IID data highlighting the need for adaptive hyperparameter optimization strategies to improve performance. In this paper, we propose a novel hierarchical federated learning solution for efficient model aggregation and hyperparameter tuning, specifically tailored to household energy prediction. The households with similar energy profiles are clustered at the edge, linked, and aggregated at the fog level, to enable effective and adaptive hyperparameter tuning. The federated model aggregation is optimized using hierarchical simulated annealing optimization to prioritize updates from the better-performing models. A genetic algorithm-based hyperparameter optimization method reduces the computational load on edge nodes by efficiently exploring different configurations and using only the most promising ones for edge nodes’ cross-validation. The evaluation results demonstrate a significant improvement in average prediction accuracy and better capturing of energy patterns compared to the federated averaging approach. The impact on network traffic among nodes across different layers is kept below 30 KB. Additionally, hyperparameter tuning reduces the size of model updates and the number of communication rounds by 30%, which is particularly beneficial when network resources are limited.

The large-scale integration of distributed renewable energy sources and electrified end uses increases the need for accurate, scalable assessment of residential demand-side flexibility. This paper presents FLEXiT, a web platform that quantifies and forecasts flexibility at aggregated (building) and disaggregated (appliance) levels. FLEXiT couples bidirectional long short-term memory models for 1–7 day demand/production forecasts with denoising autoencoders for high-resolution appliance disaggregation. A dynamic 30-day weekday baseline supports hourly upward and downward energy flexibility indicators, and a time-interval classification highlights windows for demand-response activation. Pilot deployments show FLEXiT can ingest heterogeneous datasets, visualize flexibility envelopes, and inform load shifting for aggregators and prosumers. Unifying aggregated and disaggregated forecasting in one workflow enables portfolio-scale flexibility management with appliance-level interpretability.