Ray for Large-Scale, Time-Series Energy Forecasting to Plan a More Resilient Power Grid

Kevala is building software for electrical utilities and regulators to forecast future conditions on the power grid and determine what components could be at risk as renewable energy generation and electric vehicle usage increase. This requires predicting behavior across large geographic regions covering millions of households over multiple years and simulating the complex interactions between technologies like residential solar, battery storage, and electric vehicles.

In this talk, I will discuss how we've used Ray Core, Ray Serve, and KubeRay to create a flexible and efficient architecture that distributes these forecasts across Ray actors that work semi-independently to predict behavior in specific geographic regions. Our architecture allows these forecasts to be highly configurable by our users and run on demand to process terabytes of input data and generate up to hundreds of billions of output data points. I'll walk through how we've used different features of Ray to solve challenges such as coordinating work between inter-dependent tasks, minimizing latency, and efficiently caching and transferring data throughout the workflows.