Neural surrogate-based model predictive control for compressed air energy storage

Compressed air energy storage is a promising technology for large-scale, long-duration energy storage in renewable-integrated power systems. However, accurate modeling and control of compressed air energy storage under realistic operating conditions remain challenging due to the tightly coupled thermodynamic processes involving multi-stage compression, heat exchange, cavern storage, and expansion, all of which interact under off-design conditions. This thesis develops a learning-based framework for the modeling and control of adiabatic compressed air energy storage systems under off-design operating conditions. A detailed thermodynamic model of the compressed air energy storage system is first developed to capture component-level off-design behavior, including variable isentropic efficiencies, expansion ratios, and load-dependent heat exchanger dynamics. Neural controlled differential equation models are then trained on two complementary datasets: component-level off-design performance maps and system-level sequential operational trajectories. The learned surrogates are integrated within a model predictive control framework applied to a wind-diesel-compressed air energy storage hybrid energy system representative of a remote northern community. Different neural controlled differential equation modeling formulations are introduced, including extensions of input structures to incorporate control-relevant decision variables, and evaluated across multiple dimensions: generalization to unseen operating conditions, hyperparameter sensitivity analysis, and computational efficiency benchmarking against the full thermodynamic model. A robustness analysis under additive Gaussian forecast noise confirms that the model predictive control framework maintains its fundamental dispatch character under degraded forecast quality. The developed methodology contributes to advancing the modeling, optimization, and control of compressed air energy storage systems under real-world operating conditions, supporting the broader integration of large-scale energy storage in renewable-powered grids.