Forecasting Maximum Temperatures

Next week looks hellish, with the National Weather Service currently predicting (as of 8:25 am Friday 11 June 2021) maximum temperatures for the Salt Lake City International Airport of 102˚F Monday, 103˚F Tuesday, 101˚F on Wednesday, and 100˚F on Thursday.  These are all near record highs for those dates.

How are maximum temperature forecasts produced?  The exact approach may vary depending on your source, but here are the basics.

Numerical weather prediction models form the backbone for weather forecasts.  These models solve the governing equations for the atmosphere, which are based on physical laws such as conservation of momentum, mass, and energy.  On example is the Global Forecast System, or GFS, which is the primary modeling system produced by the National Weather Service for medium-range forecasting.

Forecast models like the GFS have revolutionized weather forecasting.  In the middle of the 20th century, weather forecasting was based almost entirely on painstaking manual analysis of surface and upper level data, forecaster intuition, rules-of-thumb, and other ad hoc techniques.  Forecast models like the GFS have grounded the forecast process in science and greatly improved the accuracy of weather forecasts.  And, they continue to improve with increasing spatial resolution (enabling more detail), better accounting of physical processes related to clouds and radiation, and the assimilation of an increasing array of weather observations, especially those from satellites.

Models like the GFS, however, struggle predicting conditions very near the Earth's surface, including variables like temperature or maximum temperature.  There are a number of reasons for this, but in our part of the world, one of the biggest sources of error is the inability of models like the GFS to resolve the topography.  The GFS is comprised of grid cells that are about 13-km on a side.  As a result, does not account fully account for the topography of the Salt Lake Valley and surrounding ranges.  Since the elevations of the GFS do not match those in the real world, a GFS forecast for the Salt Lake City International Airport is strongly biased.  

This is where machine learning comes in. Around 1970, meteorologists began to train past forecasts with observations to develop techniques to produce better surface weather forecasts.  In a classic paper, Glahn and Lowry (1972) called this technique Model Output Statistics, or MOS.  MOS greatly improves upon direct model forecasts by adjusting for systematic biases and other effects.  For all intents and purposes, it was an early machine learning approach.  

The National Weather Service continues to call forecasts produced using such machine learning techniques MOS.  They also have a new system called the National Blend of Models (NBM) which uses many modeling systems to produce forecasts.  Private sector companies are also using machine learning techniques to produce forecasts, including many of those that you access on your phone.  The quality of such forecasts depends on several factors including the quality of the underlying model or models, length of the training period, and statistical techniques used.  

Where do humans enter the loop?  It depends.  Although machine learning can be quite automated, humans still are involved to some degree in the design of the approach and hopefully the testing an evaluation of the forecasts.  However, after that, most of the forecasts you get on your phone are probably generated automatically with no human intervention.  National Weather Service forecasts can, however, be adjusted by humans using software called the Graphical Forecast Editor.   A reminder here that we're talking strictly about temperature forecasts.  Humans play many essential roles in the forecast process including the issuing of watches and warnings and decision support activities.

All of this physics and statistics is wonderful, but when we have a potentially high-impact weather event, it's good to have some idea of context.  For this, I'm going to use the GFS forecast of 700-mb (about 10,000 feet above sea level) temperature.  This level is high enough that it is not strongly biased by the poor terrain representation, yet it is also strongly coupled to the surface during the summer when intense surface heating drives turbulence that strongly couples that level with the surface.  

The 0600 UTC initialized GFS forecasts 700-mb temperatures at the Salt Lake City International Airport of 17.7˚C on Monday afternoon, 19.4˚C on Tuesday afternoon, 18.0˚C on Wednesday afternoon, and 18.4˚C on Thursday afternoon.  These are exceptionally high relative to past upper-air observations.  The highest 700-mb temperature in the upper-air record at the airport is 20.2˚C on the afternoon of 13 July 2002 when Salt Lake City hit its all time record high of 107˚F.  The highest 700-mb temperature measured in June is 19.0˚C on the afternoon of 29 June 2013 when we set the June record of 105˚F (note: The prior day also hit 105˚F with a 700-mb temperature of 18.2˚C).  

Also catching my attention is that the airmass that will be in place early next week is remarkably dry and the low-level flow is southwesterly, as illustrated by the time height section below

These are also consistent with high maximum temperatures at the Salt Lake City Airport.  I took a quick look at the Euro and its numbers are close to the GFS.  On Tuesday afternoon, it has 700-mb temperatures of 19-20˚C over the Salt Lake Valley.  

We could also also take a look at an ensemble product.  Below is the NBM guidance from the National Weather Service based on many ensemble modeling systems.  Box-and-Whisker plots for maximum and minimum temperature are on the left and a graph of the median hourly temperature and dewpoint forecast is on the right.  We'll focus on the former.  The median forecast maximum temperature for Tuesday afternoon is 105˚F.  The interquartile range, or middle 50% of forecasts, lie between 103˚F and 106˚F.  The extremes look to be about 100˚F and 107˚F.  

A temperature of 105˚F would be unprecedented for so early in the season.  The earliest 104˚F was recorded on June 21, 1961 and the earliest 105˚F was recorded on June 28, 2013.  I suspect the duration of heat next week will probably make the average maximum temperature for the period the highest on record for that period.