Differentiable programming for transient power system simulations

The power systems community relies heavily on numerical simulations to understand the dynamic behavior of the system, with the ultimate goal of reliably planning and operating the system. This work proposes a method for end-to-end differentiable power system simulations, including not only the numerical solution of the differential equations, but also the solution of the power flow equations and initialization of the device models. This work builds on existing methods for computing sensitivities of the numerical sub-problems which comprise a power system simulation. The composability and flexibility of the proposed framework takes advantage of state-of-the-art numerical solvers and automatic differentiation tools while making benchmarking various methods simple within a power system context. Numerical results demonstrate downstream use cases of the computed gradients in solving optimization problems, including the training of data-driven dynamic models directly within an existing power system simulation tool. • Proposes end-to-end differentiable power system simulations. • Composable software differentiates numerical solutions of differential equations. • A case study demonstrates online training of the models in a simulation environment.