Deployable Ensembles for River Forecasting

Description:

Anticipation of rapid changes in river flow, especially floods, is a critical task for many stakeholders. While river flow rate is directly correlated to weather, these systems are difficult to model with traditional statistical methods. Neural networks are well-suited for this task, but the amount of computation required for accurate forecasting prohibits real-world deployment. We hypothesize that an ensemble of networks can outperform single network solutions while and will be easier to train and deploy. Here, we demonstrate this in practice and at scale. We present a containerized ensemble of neural network that run scheduled inference on AWS with results published on a public front end.