Generalizable Active Power Curtailment Methods for Integrating EVs and PVs in Power Grids

Active power curtailment (APC) is an essential and efficient method to prevent voltage and current limit violations in distribution grids with high electric vehicle and solar panel adoption. In previous works, reinforcement learning (RL)-based formulations have led to APC methods that can be deployed to multiple radial grids without re-training. This is achieved by reducing the action space from nodal curtailments to uniform per-feeder curtailments, lowering the amount of required training data, albeit with a potential trade-off in performance. To address the latter issue, in this paper, low-complexity APC strategies are developed that disaggregate the uniform per-feeder curtailments over the buses, while relying only on limited and coarse grid model information and preserving the generalization properties. The proposed per-bus APC approaches involve training the uniform per-feeder RL controller considering disaggregation in the environment. In addition, a heuristic generalizable APC controller that does not depend on the uniform per-feeder actions is developed. All proposed APC controllers are thoroughly evaluated and compared as well as benchmarked with respect to optimal power flow solutions. It is demonstrated that disaggregation significantly improves the performance of the uniform per-feeder RL-generated actions in terms of trading-off violations with curtailments. Importantly, the adaptability of the proposed APC controllers to unseen grids without re-training or re-tuning is showcased.