GridWarm: Towards Practical Physics-Informed ML Design and Evaluation for Power Grid
When applied to a real-world safety critical system like the power grid, general machine learning methods suffer from expensive training, non-physical solutions, and limited interpretability. To address these challenges, many recent works have explored the inclusion of grid physics (i.e., domain expertise) into their method design, primarily through inclusion of system constraints and technical limits, reducing search space and crafting latent space. Yet, there is no general framework to evaluate the practicality of these approaches in power grid tasks, and limitations exist regarding scalability, generalization, interpretability, etc. This work formalizes a new concept of physical interpretability which assesses 'how does a ML model make predictions in a physically meaningful way?' and introduces a pyramid evaluation framework that identifies a set of dimensions that a practical method should satisfy. Inspired by the framework, the paper further develops GridWarm, a novel contingency analysis warm starter for MadIoT cyberattack, based on a conditional Gaussian random field. This method serves as an instance of an ML model that can incorporate diverse domain knowledge and improve on different dimensions that the framework has identified. Experiments validate that GridWarm significantly boosts the efficiency of contingency analysis for MadIoT attack even with shallow NN architectures.
READ FULL TEXT