| الوصف: |
Effective power flow (PF) modeling critically affects the solution accuracy and computation complexity of large-scale grid optimization problems. Especially for grid optimization with varying topologies for enhanced flexibility and resilience, a tractable approximation of nonlinear AC-PF is of paramount importance. This work develops a data-driven approach to obtain piecewise linear (PWL) PF models by using the ReLU activation and an innovative neural network (NN) layer design to match the generative structure of AC-PF models like nodal power balance. Accordingly, the proposed generative NN (GenNN) PF model not only maintains the consistency among the predicted power variables but also neatly includes the topology decision variables for attaining a mixed-integer linear program (MILP) based reformulation of grid topology optimization problems. We further develop an area-partitioning based sparsification method to reduce the number of GenNN weight parameters and thus the model complexity. Thanks to our sparse GenNN, the proposed PWL-PF can achieve scalability for large-scale systems and allow for efficient solutions of AC-PF based optimal transmission switching (OTS) and restoration order problems (ROP). Numerical tests on the IEEE 118-bus and the 6716-bus synthetic Texas grid systems have demonstrated performance improvements over competing alternatives in approximating the AC-PF and accelerating topology optimization solutions. |