Researchers from ETH Zurich, including Dong Liu, Juan S. Giraldo, Peter Palensky, and Pedro P. Vergara, have developed a new approach to voltage regulation in distribution networks using distributed reinforcement learning and local smart meter data. Their work, published in the journal IEEE Transactions on Power Systems, aims to address the computational and privacy challenges of centralized reinforcement learning methods.
The team’s proposed method integrates a dynamic Thevenin equivalent model within smart meters, enabling local voltage magnitude estimation using only local data. This approach reduces the need for synchronized data collection and centralised power flow calculations, which can be computationally intensive and raise privacy concerns. By using local data, the reinforcement learning agent can be trained more efficiently and with greater respect for data privacy.
To address potential inaccuracies in the Thevenin model, the researchers introduced a voltage magnitude correction strategy. This strategy combines piecewise functions and neural networks to correct large errors and improve the precision of action adjustments. The piecewise function corrects significant errors, while the neural network mimics the grid’s sensitivity to control actions, allowing for more accurate voltage regulation.
The researchers also proposed a coordination strategy to refine local reinforcement learning agent actions online. This strategy prevents voltage magnitude violations that can occur when multiple independently trained agents take excessive actions. The coordination strategy ensures that the actions of individual agents are compatible with the overall goals of the distribution network.
Case studies on energy storage systems validated the feasibility and effectiveness of the proposed approach. The results demonstrated that the method can improve voltage regulation in distribution networks, offering a promising solution for the energy sector. The researchers suggest that their approach could be particularly useful for integrating distributed energy resources and managing the increasing complexity of modern power systems.
The research was published in the IEEE Transactions on Power Systems, a leading journal in the field of power and energy systems engineering. The work highlights the potential of distributed reinforcement learning and smart meter data to enhance the efficiency and reliability of voltage regulation in distribution networks.
This article is based on research available at arXiv.