D8.1 Trustworthy Predictive Models for Energy Demand and Production: State-of- the-Art

For this reason, the deliverabledoes not frame forecasting as a competition between model classes. It instead organises the field around practical decision criteria: consumption versus production forecasting,short-term versus long-term horizons, univariate versus multivariate inputs, one-step versus multi-step outputs, deterministic versus probabilistic forecasting, and white-box, black-box, and hybrid model families. This is the right framing for D2ET, because the project needs usable forecasting components, not isolated benchmark winners.

A central result of the review is that data requirements are as important as model choice. Forecasting performance depends not only on the algorithm, but on whether the available data capture the mechanisms that drive the system. Historical load or production measurements are essential, but they are often insufficient on their own.

For demand forecasting, relevant inputs may include temporal features, calendar effects, weather conditions, dynamic pricing, and contextual variables. In D2ET, this is directly relevant to cases such as heat-pump demand response to outdoor temperature and tariff signals. For renewable generation forecasting, exogenous weather information is often decisive, especially in short-term horizons where numerical weather prediction may be more informative than the recent history of the target series alone.

This is particularly clear for wind and solar generation. The review therefore treats data engineering as part of the forecasting method itself, not as a preliminary step. In D2ET, this is especially important because predictive models are expected to rely on datasets produced across WP1–WP5, harmonised in WP7, and later connected to automated data flows in WP9.

Another main conclusion is that the forecasting problem must be treated as multi-scale. The same project must support forecasts ranging from local assets and buildings to communities, regional infrastructures, and national system conditions. This has two implications.

First, models must be selected with attention to aggregation level, because behaviour observed at household or asset level does not transfer directly to feeder, district, or national level. Second, coherence between forecasts across scales matters, especially in a digital twin that combines multiple carriers and multiple decision contexts. For this [D2ET] [D8.1] [Deliverable Title] 3reason, D8.1 focuses primarily on electricity consumption and production while keeping the conclusions transferable to heat, gas, and hydrogen.

This is a reasonable starting point, but the broader multi-carrier logic of D2ET must remain visible throughout the report. The comparative analysis in this deliverable leads to four practical conclusions for WP8. First, simple and interpretable models should remain mandatory baselines.

They are often more stable when data are limited and are valuable when transparency is required. Second, advanced machine learning and deep learning models are justified when the problem is strongly nonlinear, high-dimensional, and data-rich, especially for renewable generation and complex demand patterns. Third, hybrid and ensemble strategies are often the most promising option, because they can combine physical intuition, statistical structure, and nonlinear learning without forcing the whole problem into one modelling paradigm.

This is particularly relevant for D2ET, where physical system logic and data- driven learning must coexist. Fourth, probabilistic forecasting should be treated as a priority rather than an optional extension whenever the forecast is used for system planning, flexibility management, or risk-sensitive decisions. Point forecasts alone are not sufficient for questions such as multi-carrier adequacy or local flexibility under uncertainty.

These conclusions align with WP8’s requirement to compare techniques by predictive accuracy, horizon suitability, computational cost, and related performance criteria. he review also shows that D2ET cannot evaluate forecasting quality only through error metrics. Since WP8 is explicitly framed around trustworthy predictive models, model selection must also consider explainability, robustness, reproducibility, and integration into a continuous operational workflow.

This is why the deliverable links forecasting methods to MLOps practices and to the later WP8 tasks on explainability and robustness. A model that is accurate but opaque, brittle under drift, difficult to update, or hard to audit is of limited value in the D2ET setting. This is not a secondary issue.

In a digital twin used for energy decision support, it is a core selection criterion. The review therefore supports a workflow in which forecasting models are benchmarked not only for predictive performance, but also for maintainability, data dependency, adaptability to streaming updates, and compatibility with transparent decision support. Finally, the review translates the general forecasting literature into the three Phase 1 D2ET scenarios.

For the heat pump versus district heating scenario, the main forecasting needs concern heating-related demand dynamics, local conditions, cost and flexibility implications, and the interaction between technical performance and public acceptability. A concrete example is the effect of cold-weather peaks on electricity demand when heat pumps are deployed at scale. For the multi-carrier guarantee-of- supply scenario, forecasting must support reliability, resilience, and coordination across electricity, gas, heat, and hydrogen, with explicit attention to uncertainty and cross-carrier interdependence.

For the deployment of energy communities scenario, the relevant forecasting problems are more decentralised and heterogeneous, involving local generation, storage, EV charging, participation patterns, and energy sharing. These scenario mappings are one of the main outputs of D8.1 because they convert a broad methodological review into concrete guidance for Tasks 8.2–8.5. In that sense, the deliverable does more than summarise the state of the art.

It defines a forecasting [D2ET] [D8.1] [Deliverable Title] 4framework for D2ET that is scenario-aware, data-aware, multi-scale, and suitable for trustworthy operation inside the national digital twin