My goal this summer was to develop a novel machine learning approach to streamflow prediction. Previous work has explored machine learning approaches to this task but they tend to use off-the-shelf methods that fail to account for the special characteristics of streamflow data. In particular, they fail to model the spatio-temporal relationships that exist within streamflow data. My plan was to develop a novel architecture using graph networks to model spatial and temporal relationships within the data. What distinguishes my work from previous work applying graph networks to spatio-temporal data is that streamflow data is heterogeneous in ways that previously studied datasets are not (i.e. different locations measure different phenomena).
With the help of Leo, I've gathered and pre-processed the data. I've also developed several simplified versions of my proposed architecture and run them on the streamflow data alone. Experiments suggest that incorporating spatio-temporal relationships into streamflow predictions is useful. I've also implemented a more complete version of my architecture that can incorporate precipitation and snow melt into predictions but this architecture is still being debugged. I hope to complete this project and submit it to a data mining conference in the Fall.