Most of the United States is well-aware of the dangers of “drinking the Kool-Aid” when it is time to form an opinion on a particular subject. However, the dangers of “eating the rainbow spaghetti” have not yet permeated the consciousness of the general public when interpreting the forecasts of hurricanes and tropical storms (tropical cyclones, or TCs). The spaghetti plot or spaghetti diagram is a visualization tool that shows the predicted paths (tracks) or wind speeds (intensities) of the numerous different TC models. Each potential TC track and intensity is shaded a different color; hence the appearance that the graphic is filled with rainbow spaghetti.
If used correctly, the spaghetti diagram can be a valuable forecasting tool. Viewing all of the potential tracks and intensities of the most realistic TC models helps scientists to understand how each model’s formulation (parameterizations, data assimilation schemes, etc.) can lead to different predicted outcomes. Additionally, the agreement or lack of agreement (commonly referred to as spread) between the models is often related to the confidence one should place in a particular forecast. If the models’ tracks and intensities are grouped together, it is often an indication that the hurricane’s future is more predictable. As a result, the spaghetti diagram can be used as a supplement to the National Hurricane Center’s (NHC) official track and intensity forecast.
When a tropical depression, tropical storm, or hurricane is present in the Atlantic or Eastern Pacific Ocean, the NHC issues an official intensity and track forecast. The intensity forecast is reported as a predicted wind speed but there are no details regarding the uncertainty in the forecast. Instead, ambitious users could look over the error statistics from past years to provide an expectation for the errors of the current storm. However, historical trends are not always the best guide for the intensity errors in individual storms, and errors often vary significantly depending on the situation. The ability to look at a spaghetti diagram and diagnose the spread of the models’ forecasts is helpful for anticipating the reliability of a particular hurricane’s intensity forecast.
Spaghetti diagrams provide a similar advantage for track forecasts. Unlike intensity forecasts, NHC’s track forecasts provide some basic uncertainty information by surrounding the predicted storm path with a forecast cone. Before each hurricane season begins, the size of the forecast cone for the year is calculated based on the NHC official forecast track errors for all storms over the past five years. The same cone is used for the whole hurricane season, no matter how confident the NHC is (see “Forecast Cone Refresher”). By evaluating the spaghetti diagram alongside the forecast cone, it is possible to foresee the situations where the cone is more reliable than others.
The 2012 track forecasts of Tropical Storm Debby are a perfect example of how useful the spaghetti diagram can be. While the NHC forecast cone was showing a developing tropical storm moving westward off the Louisiana coast, half of the model tracks were directed eastward into the panhandle of Florida. Debby eventually migrated eastward and made landfall as a weak tropical storm north of Tampa Bay, Florida. The spaghetti diagram helped reveal the particular forecast cone was less reliable than normal and that there was a possibility the storm could travel in a completely different direction than the forecast cone.
Still, the spaghetti diagram quickly loses value if evaluated by an uninformed eye. With all the cryptic model abbreviations that accompany the diagram, it is hard for the average person to develop any intuition on what models normally perform better than others. Along with the NHC official forecast (shown as OFCI on the spaghetti diagrams), there are four main types of models that are typically included in spaghetti diagrams: trajectory/statistical, statistical-dynamical, dynamical, and consensus. All of these models arrive at their predictions using different methodologies. The consensus aids are not independent; they are simply averages of other models. Some of the models you see on spaghetti plots are outlined in the table below, and a more complete list is available here.
Most spaghetti diagrams for track forecasts will include the models: “BAMS”, “BAMM”, and “BAMD”. These track models are called trajectory models and are much simpler than full dynamical or statistical-dynamical models. Trajectory models use data from dynamical models to estimate the winds at different layers of the atmosphere that are steering the TC but they do not account for the TC interacting with the surrounding atmosphere. Due to this major simplification, trajectory models should rarely be taken seriously but are included on the plots for reference. Averaged over the past five years, these models have track errors that are almost double the best performing model for a particular forecast time.
Statistical models produce track and intensity forecasts that are based solely on climatology and persistence. In other words, these models create a forecast for a TC using information on how past TCs behaved during similar times of the year at comparable locations and intensities (climatology) while also taking into account the recent movement and intensity change of the TC (persistence). Statistical models do not use any information about the atmospheric environment of the TC. As a result, statistical models are outperformed considerably by dynamical, statistical-dynamical, and consensus forecasts and should only be used as benchmarks of skill against the more complex and accurate models. The main track and intensity statistical models included on spaghetti diagrams are respectively CLP5 and SHF5. An even simpler statistical track “model” that is included on some spaghetti diagrams is XTRP (an extrapolation of the future direction of a hurricane solely based on its motion over the past 12 hours).
Statistical-dynamical models are similar to statistical models except that they also use output from the dynamical models on the environmental conditions surrounding the TC and storm-specific details to predict intensity change. The statistical-dynamical models commonly shown on intensity spaghetti diagrams are SHIP, DSHP, and LGEM. SHIP and DSHP are identical except DSHP accounts for the intensity decay of TCs over land and is therefore more accurate than SHIPS. LGEM is the best performing out of the three models. Both LGEM and DSHP are similar in skill to the dynamical models. These models are not capable of predicting rapid changes in intensity, nor are they meant to forecast intensity of weak disturbances.
Dynamical models make track and intensity forecasts by solving the equations that describe the evolution of the atmosphere. There are two main reasons why different dynamical models produce track and intensity forecasts that always differ even though they share a common goal of reproducing the physical processes of the atmosphere. First, even with the growing network of scientific instruments scattered across the globe and space, models have an imperfect picture of the current conditions in the atmosphere. This uncertainty in the current state of the atmosphere cannot be remedied; we do not have the resources to blanket every piece of the Earth and sky with instruments and measure all the necessary atmospheric parameters simultaneously. Additionally, all instruments have inherent measurement errors. Each model uniquely uses the imperfect and sometimes sparse observations available to arrive at slightly different starting points for their forecast. Secondly, even using the most cutting-edge computer systems in the world, the equations that govern the atmosphere cannot be solved for every inch of the atmosphere; it would take too long. Models have to solve equations on a 3-dimensional grid that spans the surface of the Earth and extends upward around 10 miles. Thus, even the finest resolution operational hurricane models have grid points horizontally separated by nearly 2 miles.
Scientists know that this level of detail is not sufficient; there are important physical processes happening within the grid boxes that affect the TC’s evolution. To prevent the weather that is happening at your friend’s house two miles away from being used to describe the weather at your house, modelers often use different “parameterizations”. This fancy word boils down to a variety of approximations used to extrapolate weather at larger scales (at the grid points) to smaller scales (within the grid points). The different dynamical models use a variety of grid sizes and parameterizations to capture some of TC’s small-scale processes, but these approximations ultimately lead to the models developing the TC in different ways.
The simplest dynamical model shown on spaghetti diagrams is the LBAR model, which is only a track model. Analogously to the trajectory models, the approximations used for LBAR lead to large errors and over the long-term, it is one of the worst performing models. The rest of the dynamical models depicted on spaghetti diagrams perform at a higher level. Most spaghetti diagrams include the “early models” or “early-version” of these dynamical models because they are available to NHC during the forecast cycle. These track and intensity dynamical models often include the GFDI, HWFI, and AVNI/GFSI. These models are called interpolated models (that’s the “I” on the end) because they are adjusted versions of “late models”; the previous run’s forecast is interpolated to the current time because the current run is not available yet.
The fourth class of guidance included on spaghetti diagrams is the consensus model, which is actually not a model at all. Consensus forecasts are a combination of forecasts from a collection of models, usually obtained by averaging them together. For the spaghetti diagrams of intensity forecasts, the consensus models typically included are ICON and IVCN. The consensus models for track forecasts that are normally shown are TCON, TVCE (also known as TVCN), and AEMI.
The dynamical, statistical-dynamical, consensus models, and NHC official forecast all perform at a similar level for track and intensity forecasts, while the trajectory and statistical models have significantly higher errors. Yet when someone sees one of these inferior models deviating from the rest and steering a strong hurricane into their backyard, the natural intuition is to panic. In these situations, it is important to remember which are the more skillful models.
Still, among the skillful models, some perform a little better on average than the others but there is currently no way to foresee the dominant model(s) for a particular scenario. In fact, models will seemingly have good days and bad days, good months and bad months, and even good years and bad years. That is why an informed rainbow spaghetti consumer should not focus too much on an individual noodle but instead use all of the noodles as a side dish to NHC’s forecast cone. So when staring down an approaching hurricane this season, feel free to grab a colorful bowl of spaghetti, just remember to consume with care.
– Kieran Bhatia (PhD candidate in the Department of Atmospheric Sciences)