Echo state networks for online, multi-step MPC relevant identification

Abstract

This paper presents an online nonlinear system identification algorithm for model predictive control relevant identification, based on an echo state network (ESN) trained using the recursive least squares method with a directional forgetting factor. The proposed online ESN architecture is formed by a reservoir, with fixed recurrent connections, and a reservoir readout mechanism, which is updated online to define the network output. The ESN is used to perform a multi-step prediction task, which can be used for model-predictive control purposes. The proposed online identification algorithm is evaluated in simulation using a conical tank level process and compared with the results of other approaches from the literature. The results show that the proposed algorithm is able to identify the nonlinear characteristics of the process without considering any prior information about it. In addition, the results for long-range predictions are better than the ones obtained with a model linear in its parameters and several baseline ESN models from literature, after the adaptation phase of the proposed algorithm.