Enhancing NWP Precipitation Forecasting in the NE U.S.: Post-Processing with Deep Learning
Board Location: #151
Discipline: Ecology Environmental and Earth Sciences
Subcategory: environmental science
Session: 1
Steven Aarons - CUNY Hunter College
Co-Author(s): Yanna Chen, University of Albany, Reginald Blake, CUNY New York City College of Technology, Hamidreza Norouzi, CUNY New York City College of Technology
Numerical Weather Prediction (NWP) computes the dynamics of the atmosphere and oceans to forecast the weather based on current conditions. While these models provide insights in medium-range weather forecasting, they often face limitations in accurately predicting precipitation quantity and type (p-type) as a result of regional variability and complex meteorological processes. Parameterization processes associated with precipitation are modeled by differential equations that cannot be explicitly solved, leading to errors that compound with increasing lead time. NWP forecasts often differ from observational data, especially in terms of precipitation and snow depth data. Past work examined the parameterization processes to enhance our understanding of how snowfall patterns, amounts, and p-types are influenced by the interactions between radiation, microphysics, and cumulus schemes within a case study. Even with recent advancements in NWP parametrization schemes, they cannot accurately resolve sub-grid processes and limit spatial resolution of the forecast to grid box scale.
To improve precipitation prediction, NWP model outputs can also be post-processed using deep learning (DL) methods. DL algorithms train neural networks in multiple “deep” layers through statistical techniques to identify patterns in data, leading to predictive analysis. We integrate DL techniques for the post-processing of several NWP models’ raw outputs, to enhance precipitation prediction accuracy in the Albany area. Historical weather data and real-time weather station observations are used to refine and correct raw NWP outputs. DL approaches, such as convolutional neural networks and recurrent neural networks, will be compared, combined, and evaluated to determine if they are able to enhance forecasting skill and spatial resolution in precipitation prediction. This aims to improve emergency planning, resource management, and mitigation of extreme weather impacts in the region.
Funder Acknowledgement(s): NSF REU
Faculty Advisor: Yanna Chen, ychen51@albany.edu
Role: Primary Author - Steven Aarons