This repository containts code for collecting data, training neural networks, bounding big-M, and running NN performance verification (both sequential targeted tightening (STT) and MIQP) as reported in the following preprint:
Global Performance Guarantees for Neural Network Models of AC Power Flow
This paper was submitted to the IEEE IAS for signle-blind peer review.
The models are implemented in Julia-1.10 Language, using the Flux.jl library for machine learning and the JuMP.jl library for mathematical programming. Optimization problems are solved with MOSEK (for SDPs), Gurobi (for MILPs and MIQP), and IPOPT (for NLPs). Before running the code, make sure to activate the virtual environment from Project.toml, e.g., by running
julia> ]
(@v1.10) pkg> activate .
(VerifyStrongSDP) pkg> instantiate
For any questions, please constact schevali@uvm.edu.