Deep-learning-based optimal auction design in electricity markets

Abstract

Auctions are widely used in electricity markets as a mechanism for central operators to ensure demand is satisfied in a cost-effective manner. Recently, the RegretNet framework has been proposed to tackle the optimal auction design problem with a deep learning approach. In this paper, we extend this framework to discover nearly optimal designs for electricity auctions. This is achieved by: (i) altering the neural network architecture to determine the number of units to allocate and to incorporate demand constraints; (ii) representing the information rent as part of the generator’s cost to capture individual rationality; (iii) introducing unbounded profit functions to handle capacity constrained generators; (iv) relaxing the learning problem to handle the capacity and incentive compatibility constraints; and (v) augmenting the constraints in the learning problem to handle correlated unit-costs. These extensions enable us to consider: (i) uncertain capacity and demand, possibly due to supply failures or wind–solar integration; (ii) correlated unit costs, caused by seasonal effects or shocks; and (iii) heterogeneous multi-time slot dispatch, capturing time-varying generation costs. Through experimentation we demonstrate that the method is able to recover known analytical solutions, achieving precise cost-level approximations (with errors $<1\text{\%}$) and minimal constraint violations ($\le 0.001$). Finally, we employ the method to assess the effect of renewable power integration in the Colombian wholesale electricity market. These results highlight the ability of the method to support the design of electricity markets considering the technical characteristics of the generators, the uncertainty around their capacities and costs, as well as their strategic behavior.