A Simple Neural Network to Improve GFS Temperature and Dew Point Forecasts

NOAA NCEP Global Forecast System (GFS) forecasts have persistent model biases which are in part a function of time of day, season, and climate zone. Our goal is to correct these biases by post-processing the 48-hour lead time GFS forecast. The quality of machine learning outcomes is closely tied to the quality and usefulness of the training inputs. We explore how much we can improve the air temperature and dewpoint 48-hour lead time forecasts at specific locations by training a simple neural network with a small subset of carefully selected variables that relate to weather processes rather than the “kitchen sink” approach that uses enormous data sets that often include poor quality and irrelevant data. We train the neural network using an archive of temperature and dew point forecasts and observations from selected airport weather stations around the world. Our initial set of variables includes time series of wind speed, sea level pressure, temperature, and dew point as well as time-lagged values and tendencies. We also compare the skill of the neural network to the skill of simple regression models to determine the minimal complexity required to achieve reasonable accuracy. These computationally inexpensive post-process bias corrections can potentially be valuable to many weather services that do not have the resources to run their own forecast models and rely on the free, cloud-accessible GFS output.