Causal Models for Electricity Markets and Power Grid Congestion

Machine learning models are mostly assessed based on their performance metrics, while they ideally support interventions. For power system operation and planning, these could range from utilising a different transmission line to introducing a price cap on gas to limit electricity prices. We argue that causal machine learning provides a principled framework for data-driven modelling in power grids, as it makes domain knowledge explicit, exposes modelling assumptions and quantifies the effect of interventions. We demonstrate this approach on two distinct use cases: First, we demonstrate how structural causal models for French and Spanish day-ahead markets are used to derive causally consistent explanations. Furthermore, we reproduce the effect of the Iberian gas-price cap on the electricity price in Spain, in agreement with a mechanistic market model. Second, we study the congestion of the German transmission grid. Redispatch needs are explained via the spatial distribution of renewable generation and residual loads of neighbouring countries on redispatch. Both models can be queried under conditions never observed in training and might inspire further causally consistent explanations for applications of artificial intelligence in power grids.