SpliTech 2026, the 11th International Conference on Smart and Sustainable Technologies, brought researchers and industry together from 23–26 June 2026 in the historic cities of Split and Bol (Island of Brač), Croatia. Technically co-sponsored by the IEEE Communications Society, the conference spans Energy, Smart Cities and Environment, Engineering Modelling, and eHealth. This year, CERTAIN contributed a peer-reviewed paper to the Energy track, presented by the University of Applied Sciences St. Pölten (SPU) together with its pilot and implementation partners.
The paper
The paper, “Scalable 72-Hour Generation and Demand Forecasting for Renewable Energy Communities” by Isabella Enne, Maximilian Lackner, Fabian Kovac, and Torsten Priebe, tackles a very practical obstacle to the European energy transition: operational data latency.
Austrian smart meters record 15-minute values but typically transmit them only once per day, so measurements from national data hubs reflect the system state from at least 24 hours earlier, leaving communities with complete history only up to the “day before yesterday.” To turn that into an actionable day-ahead outlook, the forecast horizon has to stretch to at least 72 hours.
The proposed framework meets this constraint with a recursive XGBoost model and an indirect forecasting strategy, predicting generation and consumption separately before computing the community balance, driven by weather forecasts rather than expensive local telemetry. A single global model configuration, wrapped in an automated MLOps pipeline (Apache Airflow + MLflow), was validated across ten diverse Austrian energy communities over a full annual cycle. Production forecasts in solar-dominated communities reached R² > 0.85 in peak months, while the study honestly maps out a “predictability gap” in communities with small-scale hydropower and consumption anomalies. The architecture is already running in production, managing forecasts for more than 25 communities.
What Sparked the Conversation
The topic drew strong interest from the audience. Attendees were especially curious about the methodology: the recursive 72-hour strategy that sidesteps the one-day latency without extra hardware, the indirect generation/consumption split, and the automated retraining pipeline that scales horizontally to dozens of communities. They were equally curious about the quality and honesty of the results across a genuinely heterogeneous, real-world portfolio rather than an idealized simulation.
A recurring thread in the discussions was where the “frontier” of predictability actually lies: the finding that winter accuracy dips are largely a low-signal statistical artifact (with mean absolute error in fact decreasing), that small commercial-dominated communities can be easier to forecast than large residential ones, and that the real hard cases are hydropower-driven profiles and holiday-period behavioral shifts. These points connected naturally to the broader SpliTech programme, from smart grids and the integration of distributed renewables, to demand-side management and load shifting, advanced metering and data acquisition, electric-vehicle and storage integration, and the financial and regulatory mechanisms that shape what data is actually available. They also resonated with the conference’s plenary vision of holistic, cross-sectoral smart energy systems, and with the sustainable-ICT theme, given the deliberate choice of a lightweight, energy-efficient model over heavier deep-learning architectures. Many conversations turned toward future possibilities: coupling forecasts with automated flexibility, and extending the approach to new communities and generation mixes.
Perhaps most notably, people responded to the practicality of what CERTAIN is building. Given the complexity of the underlying domain, visitors were pleasantly surprised by how approachable the first tool versions are to set up and use. The general sentiment was that the project addresses a real and growing pain point in the European AI landscape, and that people will be following its further development closely.
Looking Ahead
The work continues along several avenues raised in Split: bridging the “hydro-gap” by integrating river-discharge data, using seasonal-trend decomposition (STL) to separate long-term trends from short-term behavioral shifts, exploring cluster-based transfer learning so new communities can borrow model weights from structurally similar established ones, and benchmarking further gradient-boosting variants such as CatBoost and LightGBM.
Together with the EMPOWER community association and implementation partner Solutions4Energy, CERTAIN is turning these forecasts into a foundational layer for trustworthy, fair, and automated optimization in renewable energy communities.
🔗 Conference: https://splitech.org/
🔗 EMPOWER: https://beg-empower.at
🔗 Solutions4Energy: https://s4e.at