Abstract
We propose an automatic approach for manifold nonlinear autoregressive withexogenous inputs (mNARX) modeling that leverages the feature-based structure offunctional-NARX (F-NARX) modeling. This novel approach, termed mNARX+,preserves the key strength of the mNARX framework, which is its expressivityallowing it to model complex dynamical systems, while simultaneously addressinga key limitation: the heavy reliance on domain expertise to identify relevantauxiliary quantities and their causal ordering. Our method employs adata-driven, recursive algorithm that automates the construction of the mNARXmodel sequence. It operates by sequentially selecting temporal features basedon their correlation with the model prediction residuals, thereby automaticallyidentifying the most critical auxiliary quantities and the order in which theyshould be modeled. This procedure significantly reduces the need for priorsystem knowledge. We demonstrate the effectiveness of the mNARX+ algorithm ontwo case studies: a Bouc-Wen oscillator with strong hysteresis and a complexaero-servo-elastic wind turbine simulator. The results show that the algorithmprovides a systematic, data-driven method for creating accurate and stablesurrogate models for complex dynamical systems.