PowerFORM: A General-Purpose Foundation Model for Power Systems

Power systems pose multiple operational tasks over the same system state. Existing grid foundation models focus on topology generalization, typically for optimal power flow (OPF), and handle other functions by simply adding decision variables at the output. We propose PowerFORM, a general-purpose grid foundation model in which each decision variable enters where it appears in the physics, preserving their interactions across intermediate representations. PowerFORM supports both topology and cross-task transfer. It matches state-of-the-art grid-foundation-model backbones under the same data and training budget, and demonstrates cross-task transfer to transmission network reconfiguration from an OPF-pretrained model.