Forecasting II: Portland and the Polar Vortex

Physics Model Prediction vs. AI LLM predictions. Portland wrecked. The future of weather forecasting and society confronted by increasingly unstable weather.

It is now 10 days since the long threatening Polar Vortex, a huge pool of Arctic frigid air over Alberta, sank down into our region. As a city we were only partially prepared. My wife and I spent two days making last minute excursions to grocery and hardware stores laying in supplies, seeing other last minute shoppers- all of us preparing for the storm, but none of us had any idea of how bad it would be. This lack may in part be laid at the feet of weather forecasting.

Portland, Oregon as a special case

Portland metro may be the most difficult city in the country to be a weather forecaster for. It has a large elevation differential for one: at the airport almost sea level and at the top of the West Hills about 1.200 feet above that. The Cascade Range east of town runs between 5,000 and 11,700 feet at Mt. Hood. The great Columbia Gorge cuts a wind tunnel right through the Cascades connecting hot or cold continental air masses to the marine air masses to the west of town. The outlet of these often hurricane force winds is right at the eastern edge of our city. Portland is a battlefield where oceanic and continental air masses contend for mastery in a city of variegated topography. Local microclimates abound. We’re really about five cities when it comes to weather forecasting. Complicating the forecasters' task are such things as adabiatic warming and cooling, over-riding air masses and stacked inversion layers: cold at ground level, warm at 4,000 feet, cold again at 9,000 feet. It's a witch's cauldron of complexity; other cities in this country may share one or several of these factors, but none so far as I know share all- especially the wind tunnel Gorge. Forecasting here is much more difficult than in towns like say, Omaha or Wichita or Oklahoma City or Amarillo- typical Midwest towns or even in urban areas east of the cross timbers.

Our storm

On Friday, January 12th, the cold arrived with high winds. Saturday morning the transformer station three blocks from our little house, blew up with a deep boom and all the lights went out. Three seconds later power was back on, but all our digital clocks had to be reset. Saturday was a day of literally howling icy wind and blowing fine snow. Sunday a giant tree was blown over blocking four lanes of the Naito Parkway only two blocks from our house. I put on my Marmot parka, supposedly good for Himalayan expeditions, and went out to photograph it. It was like being on top of that plateau near Mt Conness in the high Sierra dead of winter (something I had done when part of a mountaineering club in college). Wind strong enough to make standing difficult and ice so strongly driven as to feel painful. This in pleasant, mild-wintered Portland! That tree in our neighborhood would only be one of thousands to fall eventually. One single house was hit by ten of them! Our neighbor's house in the West Hills was hit by one that took a 50-ton crane to remove. Our best friends were without heat or power for seven days. It was a difficult time for us and everyone in the country.

Physics-based models

Weather forecasting as we've traditionally watched on television is largely based on the use of incredibly sophisticated atmospheric physics weather models: GRAF, GFS, NAM, ECMWF, ICON, HRRR, etc. The complexity of these models is bewildering, the science is very sound, and the accuracy, while falling off with time in the forecast range, is reasonably accurate out to about ten days. In actual television forecasts the weather announcer is employing s hybrid method of physics-based models, ensembles of models, and the announcer's own historic knowledge of how local conditions effect the models' forecast. In this latter application of history, the forecaster has been anticipating a trend of the future.

In Portland the forecasters' were a little adrift. The storm was a concatenation of storms- there were more waves of ice precipitation events than originally predicted, the cold air was much more tenacious than originally predicted, the southerly flow was slower establishing itself than originally predicted. The forecasters were hampered by a format that required them to make uniform predictions for the city as a single political unit, rather than the four or five climatic areas it really is. They were also hampered by the limitations of the forecasting systems currently in use.

Enter the AI. To the rescue?

Remi Lam’s team at Google Deep Mind research have developed an LLM trained on a huge data set of historic ECMWF material covering the last forty years. This AI, called Graphcast, does not rely on physics models- it uses historic data rather than atmospheric physics equations. It is significantly more accurate and more granular in the regional area being forecast for. It also required fewer elements of input to create the output:

“For inputs, GraphCast requires just two sets of data: the state of the weather 6 hours ago, and the current state of the weather. The model then predicts the weather 6 hours in the future. This process can then be rolled forward in 6-hour increments to provide state-of-the-art forecasts up to 10 days in advance.”

Since Graphcast is already trained (at great expense) on millions of pieces of ECMWF ERA5 forecasts, much of the television forecasters’ work are done for them. Graphcast is astoundingly accurate, its code has been open sourced, and ideally it can be run on consumer computers. It is in short a revolution. It is not intended to replace physics models, which can still outperform it in certain contexts, but to augment them. We could have used it here in Portland.

Portland in the aftermath

I write this on January 24th, twelve days after the storm(s). There are innumerable trees toppled over, hundreds of damaged homes, thousands of burst pipes, public schools damaged, damaged city parks, fallen power lines, ruptured city water mains, still blocked streets. The governor has declared a state of emergency. Could this all have been avoided with more accurate forecasting conjoined with more thorough preparation and infrastructure hardening? Perhaps not.

Enter the Dragon

The wild card is climate change- the atmospheric and oceanic dynamics are rapidly changing and becoming more unpredictable, with no historical precedents for hundreds of thousands if not millions of years. Oscillations are becoming more rapid, the extrema are widening, heat and water transport mechanisms are rapidly changing. Weather forecasting models that rely on historic data sets are going to have trouble. As will local forecasters.

The future

My city, our country are sailing into a stormy sea. All we can be sure of is that the waves will be high, the sextants unreliable, of the captains and officers we can hold little trust if experience proves true. We can only pray our ship will be stoutly bound and cross-braced, crewed by seamen united rather than at war with themselves. But experience here also, provides little grounds for optimism.

In light of all this we might call our ship of the future, The Endurance II.

“I fear me.”