Safety-aware load sharing in compressor stations using constrained Bayesian optimization

Abstract

We present a unified, self-learning framework for real-time load-sharing optimization of compressor stations in parallel and series configurations. The proposed optimizer operating at the supervisory layer models the station as a data-driven steady-state mapping from shaft torque inputs to measured performance variables, thereby avoiding explicit use of compressor maps within the optimization layer. Constrained Bayesian optimization is used to learn Gaussian Process surrogate models for the economic objective and surge-related operability constraints directly from plant-response data. Safety is incorporated through measured distance-to-surge information, allowing the optimizer to favor energy-efficient operating points while accounting for uncertainty in the learned surge-margin predictions. The framework is initialized using only three operating points and is evaluated in a high-fidelity dynamic simulation environment for both parallel and series compressor stations. Across the reported simulation scenarios and initialization cases, the method consistently converged to improved load-sharing solutions while maintaining positive steady-state surge margins and without observed surge control line crossings or anti-surge controller activation. Relative to an equal-load baseline, the proposed framework achieved energy savings of up to 2.6% in the parallel configuration and 3.1% in the series configuration. These results demonstrate the potential of the proposed CBO-based load-sharing framework to improve compressor-station efficiency while respecting operability constraints in a data-driven optimization setting.