A new way to update digital models and control systems together

Researchers developed a framework to synchronize digital twins with real systems using operational data.

A new reinforcement-twinning framework helps update and synchronize digital twins of actuated systems. This method uses data from how a system is controlled during operation.

The framework uses three computational components. A Gaussian-process surrogate combines twin predictions with a learned discrepancy from measured returns. This allows Bayesian policy search to use the twin without being tied to its predictions.

Candidate policies are proposed around the policy that is optimal for the twin. Measured returns can expand this region. Physical episodes are replayed on the twin after observer-based synchronization with sparse measurements.

The framework identifies parameters by combining Bayesian optimization with gradient-based refinement. Selected information-oriented episodes use an excitation policy to improve parameter identifiability. This was demonstrated on a Burgers-flow analogue of dynamic-induction control in wind farms.

The results support reinforcement twinning for jointly adapting models and control policies. This is useful when physical interaction is limited and model discrepancy cannot be neglected.

Key facts

  • The framework uses a Gaussian-process surrogate to combine twin predictions with measured discrepancies.
  • The method was tested on a Burgers-flow analogue of dynamic-induction control in wind farms.
  • Reinforcement twinning allows for joint adaptation of models and control policies.

All sections

Edition