Analyzing deterministic and stochastic influences on the power grid frequency dynamics with explainable artificial intelligence

Power grids are essential for our society, connecting consumers and generators. Their frequency stability is impacted by supply and demand changes, including deterministic and stochastic dynamics, e.g., from market activities or fluctuating renewables. The first two Kramers–Moyal coefficients allow for a description of both the deterministic (via drift) and stochastic (via diffusion) aspects of these dynamics. Such a description and understanding could be critical to stabilizing power systems. However, how drift and diffusion differ between synchronous areas, how they vary over time, and how the generation mix influences them, remains unclear. Analyzing temporal patterns in drift and diffusion for frequency data from Australia (AUS) and Continental Europe (CE), we reveal a positive correlation between drift and diffusion. In addition, we utilize both gradient-boosted trees and neural network models to train drift and diffusion models for AUS and CE. Shapley additive explanations make these black-box models transparent and allow us to identify the total generation and load to influence the drift, while calendar features seem critical for the diffusion coefficient estimates.