Equilibrium-Consistent Reservoir Prediction and Risk-Adaptive Model Predictive Control for Nonlinear Wing-Flutter Suppression

We investigate equilibrium-consistent reservoir-computing (RC) prediction and adaptively weighted model predictive control (MPC) for a nonlinear pitch–plunge flutter benchmark at Qf=8. A conditional residual-mean network compensates for RC prediction errors, and a post-training correction enforces the known zero equilibrium. The optimizer uses deterministic corrected-mean prediction; K-nearest-neighbor (KNN) residual-error covariance adjusts state and terminal penalties, without stochastic scenario propagation. Against separately tuned fixed weights, KNN-based weighting reduced normalized integral squared error (ISE) by 1.66% in a reference Gaussian-disturbance setting with five paired realizations. Four systematic noise conditions included independent and synthetic AR(1) disturbances. An additional exploratory comparison on two new paired Gaussian noise realizations used a fully KNN-independent state-feedback weight scheduler. This simpler scheduler lowered mean normalized ISE (0.151199 versus 0.155306 for KNN) but required more input effort (1.579680 versus 1.231068) and showed slightly larger late-stage normalized root mean square error (0.008832 versus 0.008410). Thus, the comparison exposes a regulation–effort trade-off but does not establish an incremental advantage specifically attributable to KNN covariance information. The KNN trace remains an empirical weighting signal rather than calibrated uncertainty. These results are limited to offline numerical simulations, without real-time or stochastic stability guarantees.