The ABC’s of NWP

Numerical Weather Prediction or NWP, a popular abbreviation in the weather community, is a mathematical way to solve for the set of governing equations that predicts how the atmosphere will change over time provided with specified initial conditions such as temperature, moisture, wind speed, etc. In the early part of the 20th century, Vilhem Bjerknes and Lewis Fry Richardson, notable pioneers of meteorology, suggested that it would be possible to predict the weather by solving for a system of nonlinear, partial differential equations. It was not until 1946 that the first computer capable of solving such tedious mathematical integration, necessary to predict the weather, came on the scene called Electronic Numerical Integrator And Computer or ENIAC.

John von Neumann posing with the ENIAC computer (photo courtesy of NOAA).

Four years later, in April 1950, Jule Charney and John von Neumann utilizing the ENIAC produced the first weather forecast from NWP techniques. The first weather forecast was only for a 24-hour period, and it took over one day to produce a result. Therefore, the calculation process took longer than the actual weather to occur! Along the way, computing power has increased and NWP techniques have grown more robust.

In 1977, a 36-h forecast of the 500 mb had a skill score of 50 (100 is a perfect score), while the 72-h forecast of the same field has a skill score of 26. In 1990, the skill score of a 72-h forecast finally reached the level of the skill score of a 36-h forecast in 1977. Fast-forward to the current time, and the National Centers for Environmental Prediction (NCEP) has improved on the 36-h and 72-h forecast by producing skill scores of 82 and 68, respectively. Are those scores perfect? Hardly, but are they much improved from the 23 skill score produced by the 36-h forecast in 1955.

NCEP operational forecast skill score for the 36-h and 72-h forecasts at 500 mb over North America (Image courtesy of NCEP).

Numerical weather prediction is an essential part of a forecaster’s process to predict the weather accurately. While NWP can be very valuable in the overall forecast decision, they can also ruin a meteorologist’s prediction if the NWP results are improperly interpreted or produced an erroneous solution like with the Nor’easter that devastated Washington D.C. on January 25, 2000.

In the United States, the NCEP’s Environmental Modeling Center (EMC) is responsible for maintaining and improving over 20-plus numerical prediction systems, but not all of them are associated with weather prediction. Some of the other prediction systems deal with dispersion, sea surface temperature analysis, sea ice, and climate prediction. There are several NWP models within EMC’s catalogue of numerical prediction systems. One NWP model can “calculate” weather patterns out to 384 hours (16 days) out. Are the 16-day forecasts ever 100% correct...NO, but NWP models, short- or extended-range predictions, do provide forecasters with an insight into possible different atmospheric trends. NCEP NWP model outputs are free and can be downloaded by anyone since NCEP is a government-run organization. Interestingly, NWP model outputs outside of the US are not available to the general public, and a subscription fee is required for anyone wanting access to their datasets.

In this article, I will talk about several NWP models heavily used amongst weather community. My main focus is on the popular NWP models that are important for everyday forecasting for the great state of Colorado.

Global Forecast System (GFS)
GFS is a global weather forecast model that can predict weather out to 16 days and it is run four times a day (00 UTC, 06 UTC, 12 UTC, and 18 UTC). GFS is a spectral model, which is the preferred choice for global NWP models. The horizontal resolution of the GFS is T574, which translates to around 28 km between each point (calculation point) for the first 8 days of the model run; then the spatial resolution enlarges to around 70 km until the end of the model run. Does this mean the GFS can depict a weather feature that is 28 km in length in the first 192 hours of the model run? Hardly! The rule of thumb is to take the model horizontal resolution size and multiply by 5 to estimate what the minimum size a weather feature can be depicted by the model.

The vertical resolution of the GFS comprises 64 levels that are not equally distributed, but are the same number through the entire 384-h forecast time. Vertical levels are packed closer to the surface and jet stream levels for obvious weather reasons. The GFS is a system where several other numerical models are incorporated into its overall final weather prediction. The GFS is composed of an atmosphere, ocean, land/soil, and sea ice model, all working in harmony to provide an accurate depiction of future weather conditions.

North American Mesoscale (NAM) Forecast System
The NAM is a regional grid-point NWP model employed by NCEP to offer forecasters mesoscale guidance on short-term weather patterns (out to 84 hours or 3.5 days). Fundamentally, the NAM is a version of NCEP’s Non-hydrostatic Mesoscale Model (NMM) with the Weather Research and Forecasting (WRF) underpinning physics. It initializes four times a day (00 UTC, 06 UTC, 12 UTC, and 18 UTC). The horizontal resolution of the NAM over the continental US (CONUS) is 12 km with a vertical resolution of 60 layers. The NAM can generate multiple nested grids (or domains) of weather forecasts within the original forecast domain at smaller horizontal resolutions. Smaller nested grids over CONUS (4 km), Alaska (6 km), Hawaii (3 km), Puerto Rico (3 km), and significant weather events, such as fire weather (1.33 km) have finer resolutions.

The NAM 84-h 300 mb heights and winds forecast initializing on September 19, 2014 at 1200 UTC and valid for 0000 UTC September 23, 2014 (Image courtesy of NCEP).

Rapid Refresh (RAP) and High-Resolution Rapid Refresh (HRRR) Model
The RAP is a regional mesoscale NWP model as well. This model was developed at National Oceanic and Atmospheric Administration’s (NOAA) Earth System Research Laboratory's Global Systems Division in Boulder, Colorado. NCEP runs two versions of the RAP model. The first version has a horizontal resolution of 13 km along with a vertical resolution of 50 levels and forecast out to 18 hours. The second version of RAP has a significantly higher horizontal resolution of 3 km and predicts 15 hours out. This higher version of the RAP is called the HRRR. Both the RAP and HRRR are initialized hourly unlike the GFS and NAM, which are run 4 times a day. Currently, the HRRR is an experimental model, but it will become an NCEP operational model on September 30, 2014. What is so unique about the HRRR is that it uses radar-enhanced data from the RAP model to initialize hourly-updated, convection-resolving forecasts.

The RAP 18-h CAPE/CIN forecast initializing on September 19, 2014 at 1200 UTC and valid for 0600 UTC September 20, 2014 (Image courtesy of NCEP).

The HRRR 12-h composite reflectivity forecast initializing on September 19, 2014 at 1700 UTC and valid for 0500 UTC September 20, 2014 (Image courtesy of Earth System Research Laboratory).

Global Environmental Multiscale (GEM) Model
The GEM model produced by the Canadian Meteorological Centre (CMC), US equivalent of NCEP, provides numerical guidance for the short- and medium-range forecast period. GEM model comes in two varieties, a GEM Global and GEM Regional.

The GEM Global model, unlike most global NWP models, is a grid point model utilizing a horizontal variable resolution with 33 km spacing in the north-south direction at all locations and a variable east-west resolution of 50 km at 0˚ changing to 33 km spacing at 49˚N. Its vertical resolution contains 58 levels.

Even though the name suggest a regional mesoscale model, the GEM Regional is actual a global model. The horizontal grid spacing is 15 km apart and gets increasingly larger as grid points get farther away from target domain. This technique alleviates any issue with atmospheric ‘long waves’ not being handled properly during the preprocessing stage for both models’ initializations.

The GEM-Global 240-h mean SLP and 500 mb heights forecast initializing on September 19, 2014 at 0000 UTC and valid for 0000 UTC September 29, 2014 (Image courtesy of

The GEM-Regional 48-h mean SLP and 500 mb heights forecast initializing on September 19, 2014 at 0000 UTC and valid for 0000 UTC September 21, 2014 (Image courtesy of

Integrated Forecast System (IFS, Euro, European or ECMWF)
The European Centre for Medium-Range Weather Forecast operates a global forecast model called the Integrated Forecast System (IFS). Similar to the aforementioned GFS, the IFS is made up of several other numerical models to enable the most accurate weather predictions. The IFS is commonly referred to by the weather community, as the European, Euro, or ECMWF and rarely is called by its actual acronym (I will refer to the IFS as the ECMWF from here on out).

The ECMWF is the leading global medium-range NWP model with respect to forecasting verification. It is a favorite amongst forecasters and weather enthusiasts in the medium-range weather time frame. The ECMWF is prepared twice daily at 00 UTC and 12 UTC and extended out to 10 days. ECMWF is a spectral NWP model with a horizontal resolution of T1279, which is approximately 16 km. The ECMWF vertical resolution consists of 137 layers, which is more than double the GFS vertical resolution.

Prognostic weather products produced by the ECMWF are not readily available for the general public. A limited set of weather analysis is disseminated freely, while the full suite of weather “goodies” will cost private users.

The ECWMF 240-h mean SLP and 850 mb isotachs forecast initializing on September 19, 2014 at 1200 UTC and valid for 1200 UTC September 29, 2014 (Image courtesy of European Centre for Medium-Range Weather Forecast).

Ensemble Models
An ensemble forecast model is a collection of two or more forecasts, usually from the same NWP models, but with slightly different starting conditions, verifying for the same time. These ensemble forecast products are sometimes referred to as “spaghetti plots” because they have so many contours plot on one chart that it looks like a bowl of spaghetti. Each forecast within an ensemble forecast system is called a member. Ensemble members have a lower horizontal and vertical resolution than their full-version NWP model; therefore, they are not as robust when a single member is analyzed individually. Using ensemble forecasting can give forecasters added insight into a model’s confidence on the current weather trend and the possible deviation that the atmospheric flow could take.

NCEP runs two ensemble-forecast systems, operationally. Global Ensemble Forecast System (GEFS) is based on the GFS while the Short-Range Ensemble Forecast (SREF) System is made up of the NAM and WRF model runs. Both systems have 20 members and one control run which is obtained from the best initial analysis. In addition, the control run is usually what is perturbed to generate the remaining members in the ensemble. The GEFS has a global domain, whereas the SREF is a regional ensemble system.

The GEFS 72-h 500 mb heights plotted for 5400 gpm and 5820 gpm initializing on September 19, 2014 at 1200 UTC and valid for 1200 UTC September 22, 2014 (Image courtesy of NCEP)

The SREF plume for 3-hourly surface temperature over DEN initializing on September 19, 2014 at 2100 UTC and valid until 1200 UTC September 23, 2014 (Image courtesy of SPC)

The CMC has an ensemble forecast system called the Canadian Ensemble Forecast System (CEFS). CEFS is based on 21 members of the GEM to determine a range of possible weather solutions. Moreover, the GEFS and CEFS are combined to form the North American Ensemble Forecast System (NAEFS) to give forecasters a better understanding of the current and future state of the atmosphere. Finally, the European Centre for Medium-Range Weather Forecast has a medium-range ensemble forecast system as well. It is created with 51 members of the Global ECMWF model.

The ECWMF 144-h Ensemble Mean of mean SLP and and normalized standard deviation forecast initializing on September 19, 2014 at 1200 UTC and valid for 1200 UTC September 25, 2014(Image courtesy of European Centre for Medium-Range Weather Forecast).

Final Words
There are many more NWP models that the weather community uses than the operational models, which are commonly utilized by researchers, forecasters, and weather enthusiasts across North America. Each of these NWP models has their strengths and weaknesses; therefore, analyzing exclusively at one particular model could lead to tunnel vision in the general forecast process. NWP technique is intended as a complementary forecasting tool and not as a standalone decision maker for a weather prediction.

Sam Ng

Sam Ng is an Associate Professor of Meteorology at Metropolitan State University of Denver. Ng received his Ph.D. in Meteorology from Saint Louis University in 2005. Follow on Twitter @DocWX.

Denver, CO
You can unsubscribe at anytime.