What happened in the Spanish case

Generaciones Fotovoltaicas de La Mancha (GFM), a Spanish company working with distributed energy assets, developed an energy-community digital twin with the i4CAMHUB innovation hub. According to the European Digital Innovation Hubs Network account, the project combined data from photovoltaic installations, batteries and charging points in one system. Members and administrators received different views of the data, while time-series forecasting algorithms were used to predict solar generation.

This is a useful case for a small or medium-sized business operating several sites with solar generation, storage or energy-intensive equipment. The initial problem is not a lack of a clever chart. It is that actual output, consumption, battery charge and equipment status sit in separate interfaces. When readings arrive late and lack a common time base, a forecast cannot be compared reliably with reality or used to guide an action.

There is an important limit to the evidence. The EDIH account explicitly says that, at the time of publication, it could not yet provide tangible revenue or efficiency metrics; the effect would need broader adoption and operational integration to become measurable. This is a documented architecture and process example, not proof of a specific saving. Importing an ROI percentage from it would invent a result.

Telemetry first, forecast second

In the GFM account, the digital twin displays generation, consumption, self-consumption, grid injection and battery state. Administrators can see a map of sites; members see the performance of their own nodes. Alerts are intended to identify low production and consumption discrepancies. This monitoring layer can be useful even without a model: an operator can see where a problem appeared, when it started and who should inspect it.

A forecast supports a different decision: scheduling battery charging, shifting a non-critical load, or preparing to inspect a system when expected and observed output diverge. The forecast does not, by itself, diagnose the cause. A cloud, dirty panel, inverter outage and sensor fault may look similar in a single numerical series. A responsible specialist should check other signals before deciding what to do.

An IEA PVPS study of solar-irradiance forecasting shows that forecast quality depends on input data as well as method; it compares combinations of ground measurements, satellite data and numerical weather prediction. The US Department of Energy separately describes power, voltage, current and environmental measurements as the basis of PV monitoring platforms. That does not mean a small firm must buy an expensive sensor array on day one. It means model performance cannot be assessed apart from measurement quality.

A minimum setup for a Russian business

For a first pilot, choose one site and one action that can genuinely be changed. An energy manager might receive a next-day forecast and decide when to run a non-critical load. Do not start with automatic inverter control or a promise about external electricity-market settlement: accounting and control requirements depend on contracts, equipment and local rules.

A minimal flow would be:

  • the meter or inverter sends generation with a timestamp and site identifier;
  • load monitoring and, if present, storage provide consumption and battery state;
  • a weather source supplies observed and forecast conditions for the same location;
  • a validation layer marks gaps, clock errors, implausible values and outages;
  • a simple baseline forecast is compared with any more complex model on the same periods;
  • an operator sees the forecast, uncertainty range and actual deviation, then records the decision.

On-premises processing can make sense if telemetry from multiple sites must remain within a corporate network or links to remote sites are unreliable. Time series are normally handled by a specialized model or statistical method; a language model is not required. An LLM may later explain validated anomalies in plain language or retrieve a maintenance instruction, but it cannot replace measurements or the mathematical forecast. The Spanish source does not say that GFM runs its system locally: local deployment is an architectural option for businesses with their own data requirements, not a claim about the case.

Checks before buying a “digital twin”

The first risk is unsynchronized time. If one device reports every five minutes, another every fifteen, and a maintenance outage is absent from the log, a discrepancy chart may show a connection error rather than a physical event. A shared time zone, device identities, rules for missing values and a configuration-change history are needed. Also define who may view each site's commercial figures and who may change equipment settings.

The second risk is an attractive forecast without a baseline. Compare the model with a simple rule such as “tomorrow resembles a recent comparable day, adjusted for weather.” Split evaluation data chronologically, not by random rows, or future information may leak into training. Evaluate clear and cloudy days, seasons and forecast horizons separately. A single average error can hide the days when an operator most needs help.

The third risk is counting savings before an action changes. A forecast on a screen does not prove a financial benefit if the operating schedule remains unchanged. Write down which decisions are allowed, who approves them, how outcomes will be measured and what the control period is. In some processes, a reliable sensor-fault alert is worth more than a small improvement in forecast accuracy.

Economics: a model example, not GFM's outcome

Suppose a company has five sites and an engineer spends 20 minutes per day reconciling three reports for each site. Over 22 working days, that is about 36.7 hours a month. At a fully loaded cost of 800 rubles per hour, manual consolidation costs roughly 29,300 rubles a month. These are model assumptions, not data from the Spanish project and not a promise to save the whole amount. Exception review, sensor maintenance and decisions still take time. Integration, storage, backup and support also cost money.

Evaluate payback under two separate headings. The first is time saved collecting and reconciling data. The second is a changed operation made possible by an earlier alert or better forecast. The latter needs the business's own tariffs, equipment constraints and a comparison period; a generic savings percentage from another case is not credible. If a firm has a single small installation, reporting is already easy and decisions do not change, a digital-twin project may be excessive.

A next step without a long project

Collect several months of generation and consumption history for one site, its outage log and an available weather forecast. Over two weeks, put timestamps on one scale, quantify missing data, build a simple baseline and identify one decision that can be tested in shadow mode without automatic control. Then compare forecast quality and anomaly-detection speed with the current workflow. If the data do not line up, fix the measurements first. If they do line up but no decision changes, do not buy a model for its own sake.

The AI-generated illustration for VnutrII does not depict GFM's actual facilities.