| The increasing adoption of Model Predictive Control (MPC) in new applications demands computationally efficient alternatives to traditional online optimization. While neural network-based approximations of MPC policies have shown promise, ensuring the stability of the system under control remains a critical challenge. This paper investigates Echo State Networks (ESNs) as surrogate models for nonlinear MPC, leveraging their unique architectural properties to enforce control-related stability without compromising performance. Unlike fully-trained recurrent architectures such as LSTMs and GRUs, where stability guarantees often require complex constraints that degrade approximation capability, ESNs enable straightforward stability enforcement through spectral norm conditions on their fixed reservoir weights. We demonstrate this approach on the quadruple-tank process benchmark, where an ESN controller is trained to approximate an MPC policy. The optimized ESN achieves comparable tracking performance to the original MPC while reducing computational time by two orders of magnitude. Under noisy conditions, the ESN exhibits enhanced robustness due to noise injection during training, though limitations in capturing control effort penalties are observed. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.