Predicting Forecast Error of Numerical Weather Prediction Models using an LSTM

Numerical Weather Prediction (NWP) models are foundational instruments to the functioning of forecasting institutions such as the National Weather Service, academia, and the private sector. Understanding the reliability and error of NWP models is of great importance to end-users of NWP products. To augment end-user’s ability to discern the reliability of different NWP models, our LSTM model architecture aims to predict NWP model forecast error for temperature at 2 meters. Our LSTM model utilizes the New York State Mesonet (NYSM) network of 126 sites and its quality observations of atmospheric conditions as ground truth; NYSM observed 2-meter-temperature is compared with NWP predicted 2-meter-temperature to determine NWP model forecast error. New York State’s complex topography, heterogeneously mixed land-use-land-cover (LULC), and extremes in population density informed a multitude of preprocessing efforts to aid the data acquisition and manipulation for processing into the LSTM model. Our analysis aims to minimize the Mean-Squared-Error (MSE) of predicting forecast error for temperature at two meters, and to provide insight into successful preprocessing techniques. A set of experiments were designed to manipulate components of the LSTM data and architecture to improve MSE metrics for our model’s performance, and the results will be presented.