Hybrid renewable energy systems (HRES), which co-locate two or more renewable energy sources, have proven to be promising frameworks for harnessing complementarity among different renewable resources. However, the inherent uncertainty within these systems require the recourse to potential flexibility sources such as storage. This paper proposes a data-driven control scheme for scheduling the operation of a hybrid energy storage system (ESS) within a HRES comprising PV, wind and hydro generation. The objective is to maintain the generation-demand balance in real time while maximizing renewable generation intake. Multi-agent deep reinforcement learning is investigated as a decision-making tool for real-time scheduling. Its performance is compared with common state-of-the art approaches, namely model predictive control and rule-based control. The comparison is based on a set of diverse and rigorous criteria to evaluate the trade-offs of each approach. These criteria include reliability of supply, environmental impact, uncertainty handling, battery lifetime preservation, computational tractability, communication requirements, anticipative control behavior, and adaptability. The analysis highlights as well the benefits of hybrid ESS integration within a HRES. Results show that data-driven approaches can be executed with similar levels of performance as conventional control approaches. Furthermore, depending on the system characteristics and operation priorities, the selection of an appropriate scheduling scheme is a compromise between different criteria, which need to be jointly taken into account.
Published August 2022 , 27 pages
G2233.pdf (2 MB)