What makes a model connected?
A digital twin requires a defined connection between the physical system and its digital representation, plus a purpose for that connection. A static GIS view or a one-off model is not automatically a twin. Specify which states update, how often, and who acts on the results.
Four layers to align
- Asset context: stable identifiers, connectivity, elevations and operating rules.
- Observation: timestamped flows, pressures, tank levels and device states with quality flags.
- Simulation: a hydraulic representation with appropriate boundary conditions and calibration.
- Decision: a documented workflow for comparing scenarios, reviewing anomalies or recommending actions.
Where AI can help
Forecasts can provide demand boundaries; anomaly models can flag inconsistent observations; surrogate models can accelerate selected scenario searches. Keep physical feasibility checks in the hydraulic solver. Make data freshness and uncertainty visible to the operator.
Define an achievable first use
Choose one question, such as whether tomorrow’s forecast demand can be met within tank operating levels. Test the connected workflow in parallel with existing practice. A useful limited system is easier to validate than an unbounded claim of an autonomous network.
Separate the layers of a connected model
| Layer | What it must establish |
|---|---|
| Observation | Trusted timestamps, sensor identity, quality flags and known delays. |
| Asset representation | Current topology, properties, controls and operational boundaries. |
| State and forecast | A justified estimate of current conditions and future inputs. |
| Simulation | Physical consistency and suitability for the decision horizon. |
| Decision interface | Clear ownership, uncertainty, fallback and an auditable action. |
A dashboard with live data is not automatically a predictive twin. A calibrated model with occasional manual updates can still support valuable decisions. Describe the actual capabilities and update process rather than relying on the label.
Handle time and state explicitly
A tank observation received at 10:05 may describe conditions measured at 10:00. A pump-state message can arrive later than a flow reading from the same event. A model initialised from these values without time alignment may represent a state that never existed.
Store measurement time and receipt time where relevant, and define the permitted data age. Reconcile initial storage and device states before forecasting. If state estimation or data assimilation is used, retain the assumptions about sensor error and model uncertainty. Updating a state estimate is different from recalibrating permanent pipe parameters.
Test a focused prediction task
A useful first task might predict whether a reservoir remains within its accepted range over the next operating day. Inputs include current usable volume, demand forecast, available inflow, controls and relevant outages. Outputs should include a storage trajectory, uncertainty or sensitivity cases and the reason for any predicted limit breach.
Compare the prediction with observed operation over a representative shadow period. Record interventions that changed the expected schedule, otherwise a valid forecast under one planned operation may be judged against a different operation. Inspect whether improved prediction gives operators useful lead time.
Design degraded operation before adding automation
- Define what happens when an input is missing, stale or inconsistent.
- Retain a last trusted state only for a justified period, with a visible age indicator.
- Use a documented forecast or operational fallback when the connected model is unavailable.
- Separate advisory outputs from commands and enforce the established control authority.
- Record the evidence, recommendation, human decision and outcome for review.
An automatic control path requires additional assurance, permissions and operational integration. A good advisory pilot is not proof that closed-loop control is ready. Preserve existing protective controls and the ability to operate when the data connection fails.
Measure value and maintain the representation
Assess data availability, state error, forecast performance, operator usefulness and the effect on the chosen service outcome. Avoid counting a visually convincing animation as evidence of physical accuracy. A model can drift as valves change, assets are renewed or customers are rezoned.
Tie model maintenance to asset and operational change processes. Version the representation and make a model revision visible in every decision record. Follow the focused pilot framework, the forecast exercise and hydraulic validation for the supporting work.
Sources & further reading
- EPANET: water distribution modelling ↗US Environmental Protection Agency · 2020 release; living resource
External source · Checked 24 September 2026 - InfoWorks WS Pro overview ↗Autodesk · Living product documentation
External source · Checked 24 September 2026 - AI Risk Management Framework ↗National Institute of Standards and Technology · 2023 framework; 2024 generative AI profile
External source · Checked 24 September 2026
Source findings are distinguished from editorial interpretation. Apply current local criteria and project evidence when making engineering decisions.