Abstract:

As renewable energy sources and plug-in electric vehicles increasingly penetrate power distribution systems, phase imbalance becomes more prevalent, posing significant challenges for electric utilities. This paper proactively tackles this issue by implementing phase swapping sequentially at strategic time steps and locations to improve long-term operational cost savings within a given scheduling horizon. An offline data-driven phase-swapping scheduler is developed, mimicking multi-step phase-swapping optimization through an imitation learning framework via supervised learning using random forest regression. This scheduler determines the time steps and associated loads and generations requiring phase swapping. The phase swapping over the finite scheduling horizon is modeled as a multi-step mixed-integer nonlinear programming problem, utilizing a generic branch-based radial load flow model that considers both wye and delta connections for loads and generations. For an upcoming scheduling horizon, the scheduler, based on forecasted generation, load, and price profiles, first identifies the time steps and loads/generations requiring phase swapping. Subsequently, the phase swapping actions for each determined time step are solved sequentially using a mixed-integer nonlinear optimization model with phase swapping at the determined locations and with determined number of swaps. Numerical experiments are conducted on the modified IEEE 13-node test feeder to validate the effectiveness of the proposed approach.