Can Machine Learning Bring Disruptive Innovation in PLC ?

Significant innovation to power line communication (PLC) has been introduced in the last 20 years. PLC has reached high technology maturity, but moderate market penetration w.r.t. the expectations facing competition of wireless systems and the perduring (to some extent unjustifiable) criticism about its massive adoption in smart grids. The pace of innovation of PLC appears decreased, despite some recent IEEE and ITU standardization endeavors that leverage existing solutions through optimization and better customization to specific application domains. Improvements are still possible at all protocol stacks and it would be desirable they were not limited, at least in theory, by fundamental bottlenecks (which is a catchy selling point of wireless). Machine learning (ML) has powerfully penetrated many research fields including communications. It has proven to be disruptive in several domains, but the added value is still unclear when it comes to communication networks. There are two main aspects that need to be considered: a) the identification of the problems to be better solved with ML, b) the design of ML architectures tailored to communications.

In this keynote, we elaborate on both a)and b) looking at all PLC layers. Starting from the physical layer, we report recent results about learning methods for the statistical characterization of the channel response and noise in multiple conductor power lines revealing intriguing new features. Synthetic channel and noise modeling exploiting generative adversial networks (GANs) is presented. A novel methodology to learn the channel capacity (which is still unknown in PLC)exploiting the notion of autoencoders is illustrated. New modulation techniques and ML detection mechanisms are discussed. Then, moving to the upper layers, the challenge of improving routing in large PLC access networks using ML is discussed. Finally, we briefly argue on what ML can bring w.r.t. to model-based signal processing approaches for power grid diagnostics. In reference to b), we present recent work on segmented generative neural networks with data generation in the uniform probability space, which is a flexible/scalable architecture for the considered applications.