Fast Solutions in Power System Simulation through Coupling with Data-Driven Power Flow Models for Voltage Estimation

Power systems solvers are vital tools in planning, operating, and optimizing electrical distribution networks. The current generation of solvers employ computationally expensive iterative methods to compute sequential solutions. To accelerate these simulations, this paper proposes a novel method that replaces the physics-based solvers with data-driven models for many steps of the simulation. In this method, computationally inexpensive data-driven models learn from training data generated by the power flow solver and are used to predict system solutions. Clustering is used to build a separate model for each operating mode of the system. Heuristic methods are developed to choose between the model and solver at each step, managing the trade-off between error and speed. For the IEEE 123 bus test system this methodology is shown to reduce simulation time for a typical quasi-steady state time-series simulation by avoiding the solver for 86.7% of test samples, achieving a median prediction error of 0.049%.