2020 PES GM 8/4 Panel Video: Learning to Optimize Power Transmission Systems

Developments in numerical optimization and commercial solvers have allowed system operators to handle a fairly broad class of power transmission system operations, effectively and with performance guarantees. Nevertheless, energy markets of expanding footprints and participation along with the integration of renewables incur the need to solve optimal power flow (OPF) tasks at increasingly faster timescales and over uncertain input parameters. When complexity can hinder real‐time operation, learning from big data can be a viable alternative or aid to accelerate OPF solvers. This panel puts forth some recent and exciting ideas on how to exploit data and leverage machine learning tools to aid the challenging optimization tasks involved in power transmission system operation and planning. Input problem data such as renewable generation and demand can be used as scenarios to ensure probabilistic constraints in a batch or online fashion. Input data together with their OPF outputs (optimal system dispatches) can be used to train neural networks and kernels to infer the mapping implemented by an OPF solver. Importance sampling can be adopted to propagate the uncertainty of OPF inputs to probability distributions of primal (generation schedules) and/or dual (prices) variables. Historical or synthetically generated OPF solutions could be leveraged to identify binding constraints and recover approximate OPF solutions to initialize or streamline OPF solvers.