Numerical weather prediction (NWP) is vastly improving the quality of weather forecasts for both timing and location of weather events. Computing power was a limiting factor for NWP for many years, as was data collection. Now, both of these have improved, and data analysis and increasing both time and space resolution are the foci for future forecast improvements.

Types of weather models

Currently, there are two major types of models: grid point and spectral. They each have advantages and limitations, and so ensemble models try to evaluate multiple runs of these models to minimize the limitations.

Grid point models

Grid point models are similar to Finite Element Analysis (FEA) in mechanical modeling. The forecast zone is divided into a grid of discrete points. The weather conditions at each point are approximated at a specific time. For the next time increment, the weather conditions at each grid point are predicted based on the previous weather conditions at the point, as well as the adjacent points. Examples of grid point models are the North American Model (NAM) or the High Resolution Rapid Refresh (HRRR)

One of the limitations with grid point models is the edge effects, where the grid points at the edges of the model have less data to analyze, meaning the prediction is often less accurate. A fast-moving frontal boundary may go undetected or over forecasted as it leaves one model area and enters another.

This image shows a widely spaced, crude grid model of the entire planet. The resolution is far too large for predicting storms. However, increasing the resolution increases computing time and complexity. The inset shows a few of the meteorological parameters that are modeled, such as radiation, heat, momentum and so on. Source: NOAA

Spectral models

Spectral models are a different idea altogether. These treat all weather phenomena as a series of “waves,” which is extremely useful for predicting upper-level features, such as jet streams, over long periods of time. Because wave motion and interference patterns are well-studied, these models can be very accurate for large-scale and slowly changing patterns. An example of a spectral model is the Global Forecast System (GFS).

Many young meteorologists have lost trust in NWP by using a spectral model to predict local snow storms, due to its low resolution at the local level.

This NOAA WaveWatch III model is spectral. It is a natural fit to model the ocean waves, energy transfer between the ocean and atmosphere, and so on through spectral modeling. Source: NOAA

Ensemble models

Because of the limitations in different models, there is value in looking at multiple model runs to generate an “ensemble” model. Ensemble models can combine runs from different models or different run times from the same models. These are particularly useful for plotting hurricane paths, where a weighted averaging routine can be used to plot the most likely path.

Ensemble models of hurricane paths are sometimes referred to as “spaghetti models” because the paths resemble spaghetti noodles. This model was generated to predict the path of Hurricane Rita. Source: NOAA

Big data and artificial intelligence

Perhaps the most important part of NWP is determining the accuracy, or how the model is verified against reality. This is more than simply asking if it rained on a barbecue when it was supposed to be sunny.

Weather data comes from many sources. There are approved weather stations that can verify temperature, dewpoint, wind, precipitation and other data, normally collected once per hour. Some aircraft are fitted with sensors to collect upper-air data. Satellite and Doppler radar data are also integrated into the verification of NWP. Upper-air balloon soundings are conducted twice a day (0 Z and 12 Z) at most U.S. National Weather Service (NWS) offices to get a snapshot of the atmosphere.

Even with all of this data, there is potential to collect more. Numerous personal weather stations upload their data to commercial sites. State and local organizations collect some weather data, to determine if bridges and roadways will freeze. With the falling costs of electronic sensors, certain meteorological parameters (pressure, temperature and dewpoint) can be integrated into future building construction.

But how does all of this data get processed? Currently, much of the meteorological data collected is not used, as there is already too much data generated, and it is not cost-effective to review all of it. This is where developments in artificial intelligence (AI) could be extremely useful.

For example, suppose a simulated Doppler radar image is generated during a model run. AI could compare the simulated radar image to the real radar image, pixel by pixel, and look for patterns. It could develop the scoring for how accurate the data was. Perhaps humans do not care if the light green pixel of a light shower was overhead at 6:02 PM or 6:04 PM, but AI could evaluate this time shift.

NWS has the ability to launch more than two balloons a day, if the situation calls for more data. This is typically only performed ahead of landfalling hurricanes, approaching blizzards or major severe weather outbreaks. Suppose instead that AI could determine when more data was needed and could schedule additional balloon launches. While the meteorologist may not think a balloon needs to be launched on a sunny day, maybe AI picks up on a subtle feature that will impact some area downstream tomorrow. Because balloon launches are expensive and time consuming, AI may be able to weigh the cost of the data versus the value added, considering each balloon launch takes a meteorologist out of the office for half an hour to set up the balloon.

The future for meteorologists

Meteorologists will never go away, even as NWP becomes more and more accurate. They will still be required to communicate the output of the models to a consumable form for the general population. They will still be required to verify model output and ensure that the NWP “makes sense.”

The role of the meteorologist may shift more into data analysis and programming than it has in the past. It may start to look more like engineering, with the amount of data processing and coding, than the friendly television weather person. These meteorologists have always existed behind the scenes, but the focus for the next generation of meteorologists will rely less on experience and “instinct” and more on making numerical models.

NWP of the future can not only improve warning times for severe weather but may be able to tell farmers the optimal times for applying fertilizer, water or chemicals, predict the best days and times for scheduling airline flights to minimize delays, and prepare communities for lake effect snow days in advance.

Unlike other branches of science, meteorological data is being generated every day. The sooner this data can be used for NWP, the sooner water, natural resources, and ultimately human lives can be saved.