ml4ps

ml4ps (Machine Learning for Power Systems) is a Python library that facilitates the application of Machine Learning to Power Systems, with a strong emphasis on respecting the data structure.

At the core of this package is the idea that real-life power systems have a structure that varies through time (because of line disconnections, unforeseen incidents, the building of new facilities). For this reason, we believe that developing AI models that respect the graph structure of power grid data is critical to their application to real-life systems.

We provide a series of tools that were built with this concern in mind :

• a data formalism that properly describes the actual structure of the data ;

• a dataset class derived from the PyTorch data loading utility, which returns objects that exhaustively describe power grid instances ;

• a normalizer class that maps ill-distributed features into a more appropriate range of values ;

• a graph neural network implementation in JAX that respects the structure of our power grid data ;

• a post-processing class that sends the neural network output into physically meaningful orders of magnitudes ;

• an interface that allows to plug various power grid packages to our library, which are used to read power grid data, modify them and perform simulations.

Contacts

If you have in mind a use-case that would require some adjustments of the present package, feel free to contact us at laurentpagnier@math.arizona.edu or balthazar.donon@uliege.be.

Note

This project is currently under active development.

Contents

• Power Grid Data Formalism
• Dataset
• Normalization
• Neural Networks
• Post-Processing
• Power Systems Backend
• Use-cases