Predicting Forecast Errors in Numerical Weather Prediction Models Using Long-Short-Term Memory Models across New York State

Anticipating errors in numerical weather prediction (NWP) models is crucial to accurately predict the forecasts of temperature, wind speed, and precipitation for an area of interest. Forecast error is defined as the difference between an NWP model's prediction and the observed 2-meter temperature, total precipitation, and 10-meter wind speed, for a given location. To enhance end-users' ability to assess the skill of different NWP model forecasts for a specific location, our study introduces a model architecture designed to predict NWP model forecast errors. The Global Forecast System (GFS), North American Mesoscale System (NAM), and High-Resolution Rapid Refresh System (HRRR) are some of the primary numerical weather prediction (NWP) models used by forecasters to interpret the expected meteorological conditions for a given location. Leveraging the New York State Mesonet (NYSM) as ground truth — a network of 127 weather stations recording quality observations of near-surface atmospheric conditions — we have trained Long Short Term Memory (LSTM) models that can predict hour-by-hour forecast error for each NWP model, across New York State, out to 36 hours. Each LSTM model is trained on a single NYSM station, for a given forecast hour, and a specific NWP model. We will show that these LSTM models can efficiently predict forecast errors. We will also argue that the LSTM models can consistently distinguish the relationship between an NWP model's forecast and the NYSM station’s observations. This enables end-users to identify the most accurate NWP model for a given location and forecast hour, as well as provide insight into the anticipated magnitude and direction of the forecast error. Operationally, our LSTM model workflow could be used to create an interactive dashboard for end-users of NWP models, such as National Weather Service offices.