Optimal design for electricity auctions: a deep learning approach

The need for energy systems to provide affordable, sustainable, and clean energy motivates continual improvement in the design of electricity markets. However, designing strategy-proof auctions to minimize generation costs presents challenges due to technical constraints and competing goals. A recent line of work initiated by Dütting et al. (2019), has made progress in finding optimal mechanisms through deep learning. Motivated by this breakthrough, this work extends the RegretNet framework of Dütting et al. (2019) to discover nearly optimal designs for electricity auctions, incorporating capacity constraints and different assumptions regarding capacity and demand uncertainty, correlated generation costs, multi-time slot allocation, and multi-part auctions. Experiments show that the proposed approach can recover known optimal designs and find new mechanisms for electricity auctions. Real data experiments for the Colombian market illustrate the impact of wind and solar energy on improving cost stability during reduced rainfall periods. Overall, this study offers insights into market design and policy implications on competition, diversification and generation cost levels. This work hopes to contribute to providing a more flexible framework for broader applications in electricity auctions. ; Maestría