Continuous-Time Neural Surrogate Models for Simulation and Scheduling of Compressed Air Energy Storage Systems

Long-duration energy storage technologies play a critical role in supporting electric grids with growing shares of renewable generation. Among these, Compressed Air Energy Storage (CAES) offers the potential for large-scale energy shifting. Operational scheduling for CAES is computationally demanding because it relies on frequent evaluations of optimization routines and thermodynamic models that account for off-design component performance. This study introduces Neural Controlled Differential Equations (Neural CDEs) based surrogate framework designed to emulate CAES behavior. Because Neural CDEs learn state evolution as continuous trajectories, they align naturally with the dynamic processes governing CAES. Given the path-dependent nature of CAES, driven by variations in demand, renewable supply, and nonlinear physics, Neural CDEs are well-suited to represent its evolving behavior. In this study, two distinct Neural CDE surrogate models are developed: a thermodynamic CAES model that learns the internal physical behavior of the storage system, including stage-by-stage compressor, turbine, and heat exchanger pressures, temperatures, and mass flow rates, and an operational scheduling surrogate that learns optimal charge, discharge, and decision trajectories. The thermodynamic surrogate reproduces off-design behavior. Combined with the operational surrogate, the framework replicates optimal scheduling patterns, maintains physically consistent storage trajectories, and generalizes to unseen operating days, all while operating with significantly reduced computational cost. Based on these results, Neural CDEs provide a fast, physically informed, and scalable alternative to conventional CAES simulation and scheduling approaches in renewable-rich power systems.