Online Learning Enabled Hierarchical Distributionally Robust Model Predictive Control of Green-Hydrogen Microgrids under Uncertainties

Abstract

This paper proposes a novel online-learning enabled hierarchical operation framework for a green-hydrogen microgrid (MG) with a high penetration of renewable energy (RE) and hybrid energy storage. The operation framework features a bi-level multi-timescale structure which integrates a day-ahead optimal scheduling with online learning distributionally robust model predictive control (OL-DRMPC). Day-ahead scheduling optimization is employed at the high level to provide MG energy dispatch references, while the OL-DRMPC is carried out in the low level for reference-tracking under uncertainties incurred by RE and load forecast errors. To confront the imperfect knowledge of forecast-error distributions, distributionally robust conditional value-at-risk constraints are imposed on MG system states, namely the storage of battery and hydrogen, for a family of distributions called ambiguity set. By exploiting a Dirichlet process mixture model, a streaming-data-driven ambiguity set for uncertain parameters is devised over the runtime of controller and flexibly captures multimodal structure and fine-grained moment information of each mixture component. Enabled by online learning, the proposed strategy enjoys the merit of offering a reliable and intelligent MG operation. Case studies are used to demonstrate its effectiveness and superiority.