Abstract
This paper presents a transfer learning approach which enables fast andefficient adaptation of Recurrent Neural Network (RNN) models of dynamicalsystems. A nominal RNN model is first identified using available measurements.The system dynamics are then assumed to change, leading to an unacceptabledegradation of the nominal model performance on the perturbed system. To copewith the mismatch, the model is augmented with an additive correction termtrained on fresh data from the new dynamic regime. The correction term islearned through a Jacobian Feature Regression (JFR) method defined in terms ofthe features spanned by the model's Jacobian with respect to its nominalparameters. A non-parametric view of the approach is also proposed, whichextends recent work on Gaussian Process (GP) with Neural Tangent Kernel(NTK-GP) to the RNN case (RNTK-GP). This can be more efficient for very largenetworks or when only few data points are available. Implementation aspects forfast and efficient computation of the correction term, as well as the initialstate estimation for the RNN model are described. Numerical examples show theeffectiveness of the proposed methodology in presence of significant systemvariations.