invenia/metaoptopf.jl

Code for paper: Learning an Optimally Reduced Formulation of OPF through Meta-Optimization

This package defines the necessary functions to run the experiments in https://arxiv.org/abs/1911.06784.

Full details can be seen in the paper, but it has these following conceptual components:

• binding.jl: Functions for adding and removing from a JuMP certain constraints (PowerModels and JuMP):
• model.jl: Neural network training blocks (Flux). Setting and getting neural network weights for interaction with Particle Swarm Optimization
• powermodels.jl: Setting and getting grid parameters from a powermodels OPF formulation.
• evaluate.jl: Helper functions used to evaluate the computational cost of running OPF, such as timing values.

The main experiments presented in the paper are executed through ./scripts/experiment.jl. This in turn can be launched with launch.sh. These are the conceptual blocks in experiment.jl in the _main function:

• Take a model and train it 'classically' (i.e. with gradient descent). This is in the block while !do_stop(epochs, l_val). Essentially training the NN with early stopping with update!(opt, ps, gs)
• Defining a metaloss function: function _metaloss(...)
• Following this we then optimize, from this initialization, subject to the metaloss: res = optimize(_metaloss, GoCompetition.get_w(model), ParticleSwarm.... Here for example _metaloss is a function that returns the metaloss, and is passed into the optimizer. get_w gets the weights of the NN, and ParticleSwarm refers to the optimizer, etc.

Branches:

• master: DC-OPF
• ac-opf: AC-OPF