A predictive digital twin forecasts a physical system’s future state from its current state and historical behavior. Its forecast assumes that past patterns remain useful guides to what happens next. Unlike a comprehensive digital twin, it does not need to simulate multiple possible futures under changed conditions. section 1.3.3 section 1.3.4
A maintenance forecast can use different levels of modeling:
- Extrapolate a trend: a motor’s rising bearing temperature suggests it will cross a safe limit within 48 hours. section 1.3.3
- Train an ML model on months or years of operation to identify degradation patterns before failure. section 1.3.3
Predicting travel time
Google Maps combines a representation of roads, live traffic and historical trends to estimate how congestion—and therefore travel time—may change. The forecast is about the likely future journey, not just the traffic visible now. section 1.3.3
Build the representation around a decision, such as when to inspect or replace a part. Current observations alone are not enough: historical records provide the temporal context a forecast needs. Start with the simplest representation that supports the decision and its required fidelity. If the goal instead is to ask what would happen after changing an operating parameter, that calls for a comprehensive digital twin with scenario simulation and models updated using observations. section 2.7 section 1.3.4
