Lifetime Citizen Portal Access — AI Briefings, Alerts & Unlimited Follows
Purdue researchers show reinforcement‑learning VPPs can smooth prices but raise policy questions
Loading...
Summary
Andrew Lou presented a reinforcement‑learning framework for virtual power plants (VPPs) that reduced LMP volatility and shifted net‑load in Oahu simulations; he flagged regulatory and strategic risks including constraints under FERC Order 2222.
Andrew Lou of Purdue University presented research modeling virtual power plants (VPPs) that aggregate distributed resources and use reinforcement learning to set decentralized charging/discharging policies.
Lou demonstrated simulations on an Oahu test network showing that learned policies for distributed storage reduced LMP volatility and shifted the net‑load 'duck curve' as aggregations charged during abundant solar and discharged during evening peaks. "Let machines do that — don't you don't need to bid," Lou said, advocating automation of bidding decisions via algorithms that learn price beliefs and update policies.
Lou emphasized that VPPs receive LMP forecasts as input and train policies that map states (time of day, net demand, storage state) to stochastic actions. He reported consumer cost reductions in his simulated scenarios and noted the framework can scale, though he used 2‑hour time steps in the demonstrations because of computing constraints.
He also flagged policy and implementation questions: under FERC Order 2222, aggregation and nodal participation rules vary by ISO, and multi‑node aggregation remains contentious. Lou said algorithmically controlled VPPs could be price‑taking or strategic; future work must examine whether widespread automation could produce instability or market manipulation.
During Q&A, participants pressed on multi‑node aggregation, computational scale, and whether VPPs could unfairly affect congestion and locational prices; Lou noted these are open policy and research questions.
