Sparse nonlinear identification for control-oriented modeling of an organic Rankine cycle system

Accurate low-order models are essential for control and optimization of thermodynamic energy systems, yet the nonlinear and time-varying behavior of Organic Rankine Cycle (ORC) units poses a challenge for standard linear identification. This paper presents a convex sparse identification framework for nonlinear system modeling, employing a set-membership formulation to obtain compact, interpretable models with guaranteed prediction bounds. The method automatically selects the most relevant polynomial interactions from large basis functions candidates, balancing accuracy and complexity without relying on noise statistics. Experimental validation on an 11 kWel ORC test bench demonstrates excellent prediction accuracy (FIT = 88.1%) with only 46 active basis functions, outperforming linear, piecewise-linear, and multiple-model Bayesian benchmarks. The identified model preserves physical interpretability through bilinear terms representing heat-flow coupling, and its compact structure is suitable for real-time model predictive control implementation.