Independent Study Defense – Kallol Das

On Thursday, November 29 at 2:00 in the Center for Graduate Studies conference room, Kallol Das will defend his independent study “Aggregating Ensemble Weather Predictions for Rainfall Predictions”.

Abstract:
Atmospheric science extensively uses conceptual models that simulating weather. Traditionally, conceptual models are dynamical systems represented as differential equations involving `important’ variables. A meteorologist faces the challenge of choosing from the (often conflicting) predictions of different conceptual models. In this paper, we explore applying Machine Learning techniques to aggregate conceptual weather models of rainfall. Specifically, we explore the question `is it possible to form an aggregate model of existing conceptual models for predicting rainfall based only on the outputs of these models (that is more accurate at predicting rainfall)?’ Our secondary goal is to advance atmospheric science. We explore the correlation of predictive power of different conceptual models. This is a first step in categorizing the predictive power of the components (variables and parameters) of various models as a function of geographical location, time, etc. This information, in turn, helps in developing better conceptual models as well as making better aggregated predictions. We aimed to build an aggregate model for each geographical grid point, and at more than half the geographical locations, we were able to build ones that outperform existing conceptual models. We found that locations where the aggregate model was better than the existing ones were such that the best existing model had low errors, i.e., where the existing models are good, the aggregate model is better. We also discovered that locations that are geographically near each other tend to use the same group of conceptual models to build the aggregate one. We plan to analyze this to discover the commonalities in the conceptual models used with an aim to determine which model parameters are best suited for which area.