Peer-to-peer energy market clearing gets a learned speedup: a graph-network correction adds no communication and keeps the baseline's convergence guarantee
Synopsis
Clearing peer-to-peer energy markets usually relies on distributed iterations that seek a generalized Nash equilibrium, and the iteration count and communication rounds become a deployment obstacle as participants grow. A KTH team trains a graph-neural-network correction on historical clearing data and adds it to a distributed Douglas–Rachford baseline; the correction decays geometrically, so convergence is inherited from the baseline. On synthetic markets the correction cuts the iterations needed to clear and their spread, and the speedup persists when operating conditions shift to four times the training variability.
Interpretation
Peer-to-peer energy market clearing can be accelerated by historical data without adding communication rounds, with the speedup coming from a learned correction added to a distributed equilibrium-seeking iteration. Earlier distributed clearing algorithms guaranteed only asymptotic convergence to a market equilibrium, and the many iterations and the information exchanges over the trading network were treated as major obstacles to practical deployment. On synthetic markets the correction reduced the iterations needed to clear relative to the baseline Douglas–Rachford iteration and lowered the dispersion of iteration counts across instances.
Each prosumer computes the correction from local information and messages already exchanged with neighbors, so no extra communication is introduced. Replacing algorithmic updates directly with neural networks would break convergence guarantees, while a centralized structure would require collecting global information. The message-passing graph neural network shares parameters across nodes and edges, and its node and edge feature dimensions do not depend on market size or node degree, so one parameter set serves markets of different size and topology.
Convergence does not depend on the training outcome: for any trained parameters the iteration converges to a variational generalized Nash equilibrium. Learned updates usually need safeguarding mechanisms that switch between classical and learned updates, and these often require global coordination. Bounded outputs together with geometric decay make the correction summable, so the iteration can be viewed as an inexact fixed-point scheme whose convergence is inherited from the nominal Douglas–Rachford iteration.
Perspective
The demonstrated setting is synthetic instances generated from the market model of Le Cadre et al. (2020), covering trading-graph densities from the sparsest to the complete graph; neither training nor testing uses real market records.
The speedup remains substantial but degrades as operating conditions shift up to four times the training variability; under the structural change of adding or removing one prosumer, the median iteration reduction stays non-negative at every density, but a few instances on denser networks are slowed down.
