A reinforcement learning framework for optimal operation and maintenance of power grids. (1st May 2019)

• Record Type: Journal Article
• Authors: Rocchetta, R.; Bellani, L.; Compare, M.; Zio, E.; Patelli, E.
• Is Part Of: Applied energy. Volume 241(2019)
• Page Start: 291
• Page End: 301
• Publication Date: 2019-05-01
• Subjects: Reinforcement learning -- Artificial neural networks -- Prognostic and health management -- Operation and maintenance -- Power grid -- Uncertainty; Power (Mechanics) -- Periodicals; Energy conservation -- Periodicals; Energy conversion -- Periodicals; 621.042
• DOI: 10.1016/j.apenergy.2019.03.027
• Languages: English
• ISSNs: 0306-2619

Abstract

Highlights: Operations and maintenance of grids is framed as a Sequential Decision Problem. Prognostics and health management capabilities support maintenance decision-making. Q-learning with neural networks tackle high dimensional and continuous problems. A grid with prognostics and health management is considered as test case study. Comparison with Bellman's optimality highlights the doability of the method. Abstract: We develop a Reinforcement Learning framework for the optimal management of the operation and maintenance of power grids equipped with prognostics and health management capabilities. Reinforcement learning exploits the information about the health state of the grid components. Optimal actions are identified maximizing the expected profit, considering the aleatory uncertainties in the environment. To extend the applicability of the proposed approach to realistic problems with large and continuous state spaces, we use Artificial Neural Networks (ANN) tools to replace the tabular representation of the state-action value function. The non-tabular Reinforcement Learning algorithm adopting an ANN ensemble is designed and tested on the scaled-down power grid case study, which includes renewable energy sources, controllable generators, maintenance delays and prognostics and health management devices. The method strengths and weaknesses are identified by comparison to the reference Bellman's optimally. Results show good approximation capability of Q-learning with ANN, and that the proposed framework outperforms expert-based solutions to grid operation and maintenance management.