Spike Talk in Power Electronic Grids -- Leveraging Post Moore's Computing Laws

TL;DR

The physics behind Spike Talk is unraveled from the perspective of its distributed infrastructure, which aims to address the Von Neumann Bottleneck and allows adaptive and flexible control and coordination itself.

Abstract

Emerging distributed generation demands highly reliable and resilient coordinating control in microgrids. To improve on these aspects, spiking neural network is leveraged, as a grid-edge intelligence tool to establish a talkative infrastructure, Spike Talk, expediting coordination in next-generation microgrids without the need of communication at all. This paper unravels the physics behind Spike Talk from the perspective of its distributed infrastructure, which aims to address the Von Neumann Bottleneck. Relying on inferring information via power flows in tie lines, Spike Talk allows adaptive and flexible control and coordination itself, and features in synaptic plasticity facilitating online and local training functionality. Preliminary case studies are demonstrated with results, while more extensive validations are to be included as future scopes of work.

Figures (10)

• Figure 1: Development of AI models and the limit of Moore's Law. Post-Moore solutions are demanded for accommodating the new-generation AI models, with which power electronic grids involving interactive agents are facing similar challenges.
• Figure 2: Illustration on the von Neumann bottleneck: (a) mapping of the von Neumann architecture to cyber-physical power grids, and (b) distributed architecture where only local data are involved.
• Figure 3: Illustration on the spiking neuron model: (a) equivalent RC circuit, and (b) physics of spike generation with threshold $V_\mathrm{th}$.
• Figure 4: Going beyond von-Neumann computing architecture for exchange of information between neurons: (a) Binary-activated NN trained using multilayer perceptron policy with iterative learning based weight updates, (b) spiking neural networks using synaptic plasticity allows online learning using the spike-based events from power flows at each node.
• Figure 5: A system-level view of the connectivity of local neuromorphic networks. Local neural networks are coupled through the system admittance matrix $\mathbf{[Y]}_{n\times n}$.