Power Grid Behavioral Patterns and Risks of Generalization in Applied Machine Learning

The computation burden of solving nonlinear optimization Recent years have seen a rich literature of data-driven approaches problems in operation and planning has motivated the designed for power grid applications. However, development of data-driven alternatives to state estimation insufficient consideration of domain knowledge can impose (SE) [7] [19], power flow (PF) analysis [4][14], optimal power a high risk to the practicality of the methods. Specifically, flow (OPF) [2] [9][6], as well as data-driven warm starters ignoring the grid-specific spatiotemporal patterns (in load, to collaborate with physical solvers [12][20], etc. generation, and topology, etc.) can lead to outputting infeasible, Despite their popularity in recent years, people have long unrealizable, or completely meaningless predictions on been aware of the risks of machine learning (ML) tools regarding new inputs. To address this concern, this paper investigates their impracticality[11] under realistic power grid real-world operational data to provide insights into power conditions. The risks come from the "missing of physics" in grid behavioral patterns, including the time-varying topology, general ML methods. Specifically, the transient system dynamics, load, and generation, as well as the spatial differences changing topology, and varying supply and demand (in peak hours, diverse styles) between individual loads and are physical reasons behind the temporal grid evolution.