MESOSCALE NWP MODELS.pdf

27 CHAPTER 5 MESOSCALE NWP MODELS Index:PKC / NWP-P.5 Introduction 1. A mesoscale Numerical Weather Prediction (NWP) model is a model with sufficientl...

27 CHAPTER 5 MESOSCALE NWP MODELS Index:PKC / NWP-P.5 Introduction 1. A mesoscale Numerical Weather Prediction (NWP) model is a model with sufficiently high horizontal and vertical resolution to forecast mesoscale weather phenomena. These phenomena are often forced by topography or coastlines or are related to convection. They present some of our most difficult forecasting challenges. Most severe weather occurs at the mesoscale, including tornadoes and Mesoscale Convective Systems (MCSs). Visibility, turbulence, sensible weather and sea state can vary enormously over just a few kilometers and have a tremendous impact on operations. On the job, we frequently depend on guidance from mesoscale models, particularly in tactical situations where real-time weather observations are sparse or nonexistent. As a result, an understanding of how mesoscale models work can help us use these models more effectively and improve our forecasts. 2. Physical laws of motion and conservation of energy govern the evolution of the atmosphere. These laws can be expressed by a series of complex mathematical equations that make up the core of what we call numerical weather prediction. We call these equations forecast or prognostic equations, because they predict what will happen in the future. The variables in the equations represent different aspects of weather (for example, wind, pressure, etc.). Since these equations govern how the variables change with time, if we know the initial condition of the atmosphere, we can solve the equations for a later time and obtain new values of those variables. In a nutshell, this is how numerical models make a forecast. 3. Early operational weather models could resolve only synoptic scale features. As computational resources have increased, it has become feasible to model and predict mesoscale weather phenomena in a timely manner. Mesoscale models are now an important and growing subset of numerical weather prediction tools. Predicting mesoscale weather features requires much finer model resolution as well as some other significant changes to traditional synoptic and global modeling approaches. Resolution as used here is the grid spacing or wave number that represents the average area surrounding each grid point in a grid-point model or the number of waves used in a spectral model. 4. You are probably already familiar with many mesoscale models, including run by NCMRWF etc. These models are run at varying horizontal resolutions with grid spacing typically less than 10 km. As a result of their relatively fine resolution, these models are usually run over a limited area and consequently require info on the boundaries of their domains. If they were run globally, it would take so long to complete a model run that the results would no longer be useful for the short term forecast. This aspect is rapidly evolving with the increase in computational power available to organisations and operational centers running the NWP model.  28 5. The importance of running and using mesoscale model output depends on the scale of phenomena we are interested in forecasting. When comparing global and mesoscale models, there is generally little benefit gained from mesoscale model output in situations where the weather is either dominated by synoptic-scale features or characterised by interior regions with large stable air masses and subdued topography. 6. However, in regions where the environmental conditions change rapidly in time and space, mesoscale models can provide extra information that helps forecasters make the best forecast possible. These models are able to provide great detail and often accurately represent the intensity of smaller-scale weather phenomena that interest forecasters. Additionally, mesoscale models often produce superior forecasts in coastal and mountainous regions when compared to traditional larger-scale models. They do this by taking advantage of high-resolution topography datasets and detailed sea surface temperature information when they are available. Horizontal Resolution 7. The horizontal resolution of an NWP model is directly related to the size of a weather feature it can simulate. The higher the resolution, the smaller the feature the model can successfully depict. This resolution is related to either the spacing between grid points for grid-point models or the number of waves used to represent weather data for spectral models. Operationally, COAMPS, NAM, MM5 and WRF are all grid-point models. 8. So what grid size is necessary to resolve any particular weather feature? It typically takes at least five grid points to define a feature in a grid point model. The smallest weather feature that can be depicted, even within short range forecasts, span five to seven grid points. In that case, 10-km grid spacing means that the model cannot predict features smaller than about 50 km in size! In hilly or mountainous terrain, even grid spacing of 5 km may miss a topographically forced feature. 9. As an example, to capture typical wind eddies in Monterey Bay one would need a grid spacing on the order of a few km. However, to capture the essence of the flow in the much larger southern California bight region, a grid spacing of a few tens of km might do the trick. Of course, if you are interested in wind eddies in the lee of Catalina Island, then you're back to needing the same higher resolution grid spacing that you needed in Monterey Bay. 10. Here are examples of COAMPS predictions for 1000-mb winds over Monterey Bay at two different resolutions: 27 km and 3 km. The prognoses are for 2 PM local time on an August afternoon. Note the general northwesterly flow in the offshore region. The 3-km plot clearly shows the existence of the sea breeze, represented by the onshore flow around the Bay, including southerly winds near Santa Cruz. This 3- km simulation captures the sea breeze adequately, while the 27-km simulation fails. The 27-km simulation shows northwesterly flow along the coast, unperturbed by sea breeze processes. A lack of model resolution in the 27-km simulation was the primary cause of this discrepancy. This example helps illustrate how resolution limits what a numerical weather prediction, or NWP, model can predict. Remember, both  29 the 27-km and 3-km runs of COAMPS are considered mesoscale models, despite nearly an order of magnitude difference in their resolution! 11. Increasing resolution places increasing demands on computing resources because the model must calculate values for more grid points. As illustrated here, when we divide the distance between model grid points by three, the number of grid points over the same area increases 9-fold. Not only that, but as we decrease grid- point spacing, we typically decrease the length of time between intermediate forecast steps. As a result, additional intermediate forecast steps are required to produce the same length forecast! Nonetheless, higher resolution is worth the additional computational demands for many reasons, including improv represent terrain. This, in turn, affects how accurately the model can predict terrain- induced or terrain-enhanced meteorological phenomena. Vertical Resolution 12. Just as sufficient horizontal resolution is necessary to depict different atmospheric phenomena, NWP models must also be designed with adequate vertical resolution to forecast the vertical structure and effects of a variety of meteorological events. It is interesting to note that the ratio of the horizontal and vertical resolutions must be consistent with the tilt of the weather phenomena of interest. If consistency is not maintained, model forecasts with high resolution in only one dimension can actually be worse than forecasts with lower resolution! 13. When determining a model's optimal horizontal resolution, we need sufficient detail to ensure that similar weather events are forecast equally well at almost any location within the domain of the model run. This means that we use uniform grid spacing. However, when det take advantage of the fact that certain atmospheric processes are usually confined to specific vertical regions of the atmosphere. Accordingly, modelers try to place the highest vertical resolution where it is needed most. For example, vertical resolution must be quite fine (on the order of a few millibars) near the earth's surface. This allows the model to capture the transfer of heat and moisture into the boundary layer produced by daytime surface heating. This same detailed resolution is not necessary in the middle troposphere (~ 600 to 300 mb), although an increase in resolution is necessary near and below the tropopause to predict the jet stream accurately. 14. Different numerical models use a variety of vertical coordinate types to represent atmospheric layers. Each has its own advantage as well as its own limitations. It is important to understand how the vertical resolution characteristics of ability to represent weather features. Many models including the GFS, COAMPS and AFWA WRF represent vertically stacked horizontal layers using sigma ( ). In a sigma coordinate system, the bottom and top are defined to be those levels where vertical motions are negligible. pressure value where vertical motion is assumed to be negligible (0.0). Thus, near the surface, sigma layers closely mimic terrain, while aloft, layers with low sigma values flatten out and become nearly horizontal.  30 15. The use of sigma coordinates, rather than pressure or height, avoids the complication created when a surface of constant pressure or height intersects the ground. In this example, a 920-mb surface (and every other surface below 920 mb) would intersect the mountain, while a 0.92-sigma layer follows terrain. Note also how the relief diminishes on progressively higher sigma layers. 16. Because the lowest layers in a sigma coordinate system mimic topography, modelers can easily increase vertical resolution near the surface by increasing the number of layers there. Increased resolution in the lower atmosphere enables models to better define boundary-layer processes and features that contribute significantly to sensible weather elements, such as low-level winds, turbulence, temperature and stability. Sigma coordinates do have some disadvantages, however. They can sometimes lead to unrealistically strong and vertically exaggerated wave forecasts in the lee of mountain ranges. They may also fail to adequately capture terrain blocking in cases where an inversion lies below the mountain top. Hydrostatic vs. Non-hydrostatic Models 17. Most grid point models and all spectral models in the current operational NWP suites are hydrostatic. In contrast, many mesoscale models, such as COAMPS and the AFWA WRF, are non-hydrostatic. Hydrostatic models assume hydrostatic equilibrium, in which the downward weight of the atmosphere balances the upward- directed pressure gradient force. 18. This hydrostatic assumption is valid for synoptic- and global-scale systems and for some mesoscale phenomena. Non-hydrostatic processes and their effects become important when the length of a feature is approximately equal to its height. Since the heights of most weather phenomena are limited by the height of the troposphere, this becomes an issue for features approximately 10 km and less in size. Important weather examples with significant non-hydrostatic processes include convective storms, gust fronts and other convergence lines and gravity waves, including mountain waves and turbulence. 19. For numerical weather prediction, non-hydrostatic models include equations for vertical motion that hydrostatic models lack. As a result, non-hydrostatic models directly forecast weather resulting from vertical motion due to buoyancy changes and other vertical accelerations. In contrast, hydrostatic models can only infer the weather phenomena resulting from such vertical motions. 20. High-resolution non-hydrostatic models can somewhat realistically forecast changes in atmospheric buoyancy and the associated potential for convection. To accomplish this, they include an additional forecast equation that accounts for vertical accelerations and vertical motions directly, rather than inferring the vertical motion based on the horizontal convergence and divergence. Conceptually, the change in the vertical motion in a model grid box can be calculated as equal to the changes caused by both advection and buoyancy processes minus the non- hydrostatic vertical pressure gradient force and the drag caused by precipitation. To calculate vertical motions and buoyancy properly, non-hydrostatic models must include a great deal of detail about cloud and precipitation processes in their  31 temperature and moisture forecast equations. But this detail also introduces great potential for error. Consequently, non-hydrostatic models can be very sensitive to small differences in atmospheric structure. 21. Because non-hydrostatic models have to solve an additional prognostic equation for vertical motion, the sensible choice for vertical coordinates is limited and competes with the desire for finer scale horizontal and vertical resolution. As a result, most non-hydrostatic models use a vertical coordinate based on height (z). Only a few use pressure-based coordinates (p) and none use isentropic (theta) coordinates. COAMPS uses a sigma coordinate based on height, z. 22. In practice, most non-hydrostatic forecast models sacrifice vertical resolution in order to run the models in real time at high horizontal resolution. This is not a serious problem when forecasting deep convection, but reducing vertical resolution include boundary-layer structure, thin sheets of moisture drawn into sloping baroclinic zones and/or the detailed structure of the tropopause. Fortunately, as computing power increases, vertical resolution improves and this trade-off diminishes. Impact of Parameterization 23. NWP models cannot resolve features and/or processes that occur within the confines of a single grid box. Thus, even mesoscale models cannot resolve local flows, swirls, or obstacles. This example shows complex flow where turbulent eddies are created around obstacles. Friction is larger near tall trees and buildings than it is over open areas. We cannot realistically expect weather models to resolve features at this scale, no matter how high the resolution. As a result, they must account for the total effect of these obstacles and surfaces on the flow with a single number that represents friction within the grid box. 24. The method of accounting for such effects, without directly calculating them, is called parameterization. Another way to think of parameterization is modeling the effects of a process (emulation) rather than modeling the process itself (simulation). 25. This image depicts some of the many near-surface physical processes that are typically parameterized. Cloud processes are parameterized as well. The effects of these processes must be parameterized in a model for three main reasons:- (a) Computers are not yet powerful enough to directly treat them because the phenomena are either too small or too complex to be resolved numerically. (b) The processes are often not understood well enough to be represented by an equation. (c) The effects profoundly impact model fields and are crucial to creating realistic forecasts.  32 26. Many parameterization schemes exist for emulating the significant impacts of convective processes in NWP models. Even mesoscale models with high-resolution inner nests that can model convection without the use of parameterization schemes usually have to employ convective parameterizations scheme in their outer nests. For example, the Kain-Fritsch scheme is used in all the outer nests of the COAMPS model when grid spacing exceeds a few kilometers. It is physically realistic in many ways, including its handling of the moisture exchange between the storm and its nearby environment. Of all the current convective schemes (c. 2002) it has the most realistic treatment of capping inversions, convection initiation and downdrafts. Its drawbacks include (1) the tendency to leave saturated layers that are unrealistically deep in post-convection soundings and (2) significantly longer run times than simpler schemes. Note that the convective parameterization schemes were designed to reduce atmospheric instability in models. The prediction of precipitation is actually just a by-product of the way in which the scheme does this. Consequently, this scheme may not predict the location and timing of convective precipitation as well as forecasters should expect. 27. Even in the very high-resolution innermost nests of mesoscale models, there are still other significant meteorological phenomena that can dramatically impact model forecasts and so they must be parameterized. This example shows water vapor condensing, leading to droplet collision and coalescence and a release of latent heat. This all occurs in a tiny area within a 1-km grid box and must be parameterized. The use of parameterization schemes typically has its greatest impact on predictions of sensible weather at the surface. This results from the fact that all sensible weather elements at the surface are inherently mesoscale, or even microscale and are generally not accounted for by the analysis of the synoptic situation. Consequently, simple synoptic forecast rules-ofthumb become less useful when using mesoscale models. This, in turn, requires forecasters to apply an understanding of physical processes on a case-by-case basis when forecasting mesoscale weather events. 28. Problems associated with using parameterizations can result from the increasing complexity of parameterizations and from interactions between parameterization schemes. Unfortunately, forecast errors created by the interaction of parameterization schemes are more difficult to trace than errors resulting from a single scheme. Boundary Conditions and Initialization 29. The boundary conditions of a mesoscale, limited-area model, or LAM, provide yet another consideration in determining its ability to properly predict mesoscale weather features. These boundary conditions are the meteorological conditions, represented mathematically, at the edges of the area of the model run. Calculated model fields must be consistent with these boundary conditions. 30. Those familiar with the WRF model are aware of decreasing grid-point spacing via the triple-nested products. But how do these mesoscale models get their initial or boundary conditions at the edges of their domain? It turns out that most mesoscale models rely on comparatively coarse-resolution global or regional models with substantially larger grid spacing to provide them their initial and boundary  33 conditions. Of course, a global model with the same resolution and physics as a mesoscale model is superior to a version covering a limited domain because there are no boundary condition problems. But, as stated before, global mesoscale models are not computationally practical for short-term forecasts at the present time. For now, most mesoscale models can only be as good as the information fed into them from the larger scale model output. Consequently, use of any mesoscale LAM should start with validation of the synoptic forecast used for the LAM domain boundaries. Nearly all significant weather features can be affected by refraction or redirection of atmospheric waves as they pass through a limited-area model's boundaries into an inner nested grid. Notable changes can occur in precipitation fields, temperature fields, the jet stream pattern, the vertical motion field and in the intensity and placement of surface lows and fronts. 31. In addition to boundary condition problems, LAMs may have problems when - -up refers to starting the vertical motions and divergent circulations in the model and getting them up to full strength. When the model run initially starts, these circulations are often too weak or are not coherent. This occurs if the data for the initial boundary conditions comes solely from observations with no model used to build in dynamic consistency between mass and winds. When that occurs, it takes some time to get model output cranked up to full amplitude. Also, during this time, the model may generate many errant gravity waves as it adjusts toward the model's dynamic balance. 32. The spin- incorporate data, such as surface observations and soundings, over a long time to help create the analysis. The data assimilation system merges observations with the model run in such a way as to preserve the ongoing circulation. This way, the resulting analysis will not exactly match surface observations, but will reflect an ongoing evolution of conditions just prior to the analysis. 33. In contrast, a cold start typically uses the analysis from some other source, such as a global model run with a coarser grid, to start the mesoscale model running. The advantage of a warm start is that you don't have as much spin-up problem. The model has been running, creating vertical circulations associated with fronts and jet streaks, latent heating associated with precipitation and so on and those fields are getting worked into the analysis. A cold start requires the model to gradually build up divergent circulations, vertical motion and precipitation, hence the term "spin-up." The model may also respond with some oscillations early in the forecast period as the model adjusts to initial conditions not consistent with its numerics and physics. 34. It sounds like a warm start should always be superior, right? Not so fast! In a model forecast that forms the basis of the analysis. If there are no observations at the location of a small-scale spurious feature in this forecast, then the assimilation system leaves the feature intact. These spurious features can pose a serious problem. For example, at the synoptic scale, think of a storm tracking across the Pacific Ocean. The operational model runs with a data assimilation cycle are warm  34 starts. However, large swaths of the Pacific are data voids. With no data to correct an errant forecast, a spurious feature in the model forecast goes into the analysis for the next model cycle. As a result, the forecast model incestuously propagates its own errors, growing worse and worse as the storm tracks through the data void. 35. With high-resolution models, small-scale features may quickly develop in the forecast. Some of these features, although realistic looking, may be completely fictitious. Because high-resolution models are run over limited areas, they are likely to encounter data voids, especially outside the coterminous U.S. This increases the chances that a warm start analysis will contain these fictitious features, which will, in turn, potentially bust the ensuing forecast. When you use a "cold start," you are not incorporating a forecast into the analysis, so there are no spurious predictions to correct and this problem disappears. 36. Overall, however, a warm start will typically produce a better forecast, except for the spurious feature issue. The biggest difference will occur during the first few hours, decreasing exponentially thereafter. Noticeable differences may persist 6 hours into the forecast period, but generally, they should be negligible after 12 hours. Of course, these are generalizations and in some sensitive situations you do see large differences throughout the forecast, but that tends to be the exception. 37. So why do we use cold starts? The main reason is that running a data assimilation cycle with a highresolution model consumes far more computer resources than most anyone has available. So computationally, a warm start is simply not always feasible. Using Mesoscale Models 38. Happily, high-resolution, non-hydrostatic models are able to predict the details of mesoscale phenomena such as mesoscale convective systems. The model forecasts -edge gust front and associated temperature change, the thick anvil and its effect on surface temperature and the trailing mesohigh and its impact on surface winds. These details often look very realistic, but unfortunately the forecast of convective initiation is still subject to considerable error! 39. As shown in this sea breeze forecast example, mesoscale detail is generally most reliably predicted when forced by topography or coastlines. Forecasters should be aware that the detailed forecasts generated by mesoscale models are often best used as guidelines for what may occur, with the knowledge that the location and timing may have considerable error. This is especially true in the absence of orographic forcing and becomes more likely as the forecast time increases. 40. To determine if a mesoscale model forecast is on target or not, the first thing a -term forecast to actual observations. It is also useful to review a series of previous forecasts to determine how accurate the model has been recently and to determine how the most recent forecast has or has not departed from the trend. It is always important to look for strange things in the forec by comparing model fields to observations for synoptic features and their trends. In  35 an operational setting, there is typically not enough time or data to do this comprehensively, particularly for mesoscale features. 41. This figure shows an observed sounding recorded at Denver, Colorado overlain with a corresponding initial sounding from the 12-km meso-Eta model. The fields generally agree, indicating that the model initial fields are reasonably on target. However, there are some subtle differences that should alert the forecaster to keep an eye on this situation. (1) The model is not capturing low-level upslope (easterly) winds within the shallow cold air mass, (2) the observed temperatures at the lowest levels are much colder than those in the model analysis and (3) a shallow layer of very dry air between 400 and 500 mb is not captured well by the model. 42. In some instances, a model run that initialized with inaccurate boundary conditions may still prove useful. Imagine a case where the analysis missed a short wave advancing into the region of the LAM. For a short period of time, the model may still provide a useful forecast for regions that lie away from the rogue short wave. This is particularly true for predictions of topographically- or coastally-forced weather events. Once the rogue short wave influences the weather in the region, the model forecast will be a bust. But until that time, the mesoscale model may do a much better job than a global model of predicting the weather, due to its finer resolution. This type of situation may occur whether the model fails to detect a real exists. In either case, the lesso that initialized poorly. Depending on your location and the nature of your particular forecast problem, the model forecast may still provide useful guidance. 43. It can be difficult to discern the wheat from chaff in mesoscale model output. Another good example of using mesoscale model intelligently is in the interpretation of model forecasts of mountain waves. Even at mesoscale resolutions, the tendency is for the fine-scale details of the mountain wave to be smoothed out over a larger scale than real data would show. However, the fact that mountain waves appear at all in a forecast product (smoothed or not) should alert the forecaster to the possibility of turbulence and strong winds. 44. Conversely, mesoscale model output can look much noisier than output from a model with a coarser grid spacing. Compare the two COAMPS forecasts above. Notice how the 9-km model forecast has many more squiggles in the MSLP contour lines than the 27-km model forecast. This results from both the finer resolution and the fact that the output from the 9-km model has not been filtered as much to smooth the lines. If you disregard the squiggles, the two plots are remarkably similar. However, it is important to note that the squiggles in the 9-km forecast are real and not just artifacts of the model. If you had a plot of true MSLP over the Appalachians, based on a high density of observations, the field would be much more squiggly than the 9-km forecast. In fact, it would look like a complete mess unless some filtering was applied! 45. It is important for users of model output to know that in general all numerical weather prediction models are least skilled at forecasting precipitation. In this regard, most mesoscale models are better than largerscale models because they can run sophisticated schemes that directly predict precipitation and clouds. They also  36 account for internal cloud processes. These complex cloud schemes forecast hydrometeor types that are not forecast by models with simpler schemes, including raindrops, graupel and snow. 46. The strengths of these schemes are many and include:- (a) More realistic prediction of relative humidity fields. (b) Direct prediction of each precipitation type (and amount) reaching the surface. (c) Direct prediction of cooling from evaporating and/or melting precipitation particles. (d) Direct prediction of various ice crystal habits and their effects. (e) Allowing for precipitation to occur over time (rather than in just one model time step) (f) More realistic prediction of snow blowing downwind from its generation region. (g) The ability to directly forecast aircraft icing and precipitation type based on the existence of supercooled water. (h) Overall better depiction of convective systems and cirrus anvils. This is important because of the effect of ice clouds on radiation in the atmosphere. 47. Of course there are limitations to these schemes as well. First, they can become prohibitively expensive to implement. This is perhaps their principal drawback. Second, the precipitation scheme has to "spin-up" until equilibrium is reached between the hydrometeors and the forecast moisture, temperature and wind fields. This spin up typically results in the under prediction of clouds and precipitation early in the forecast cycle. And third, these schemes can be very sensitive to poor initial analyses and limited observations. As a result, they may inaccurately depict key atmospheric structures. FAQs About Mesoscale Models Question 1. Are mesoscale models better at the surface than aloft (e.g., can they forecast the sea breeze better than flight level winds)? This depends on the duration of the forecast, the quality of the initial conditions, etc. While a mesoscale model may do well predicting a sea breeze on the first day, it may be much worse by the second day. It also depends on circumstances. For instance, if the sea breeze is disrupted by convection, the model will have trouble because of difficulties getting the  37 convection right. Under quiescent conditions, because the sea breeze is forced by the land-sea temperature contrast with diurnal differential heating, as long as the synoptic low-level flow is not very badly forecast, the sea breeze will be among the most reliable model features, though its inland penetration in the model will be only as accurate as model resolution can buy. If the forcing is topographic in nature, the higher the resolution, the better the phenomena will be forecast/depicted, as long as the larger-scale conditions are right. There are issues around whether what you see at mesoscale is what you get. Realistic looking depictions do not mean accurate depictions. Mesoscale models will show details that a global model could never depict, but the forecast may be a total bust. Also limitations in what the model physics can do will limit what you see in a mesoscale model; think about how convective parameterization is handled, for example and even how cloud resolving models (mis)handle convective development! Question 2. How do I determine if the current run of the model initialized properly? To determine if a mesoscale model forecast is on target or not, the first thing a -term forecast to actual observations. It is also useful to review a series of previous forecasts to determine how accurate the model has been recently and to determine how the most recent forecast has or has not departed from the trend. It is always important to look for strange thing by comparing model fields to observations for synoptic features and their trends. In an operational setting, there is typically not enough time or data to do this comprehensively, particularly for mesoscale features. Forecasters should be aware that the detailed forecasts generated by mesoscale models are often best used as guidelines for what may occur, with the knowledge that the location and timing may have considerable error. This is especially true in the absence of orographic forcing and becomes more likely as the forecast time increases. Question 3. Does the accuracy of mesoscale model predictions decline more rapidly than that of global models? Yes and for 2 different reasons: First, a mesoscale model usually only covers a limited area. Over time, information from outside this area creeps in through the boundary conditions, approximately as fast as the wind blows. If you had a high-resolution global model, this issue would not exist. But for a limited-area model, the prediction quality declines to the quality of the information feeding the boundary conditions. For example, the limited domain NAM model uses boundary conditions from the GFS model initialized 6 hours earlier. If the NAM were run out long enough for the boundary conditions to influence the whole domain, its skill level would drop to approximately the level of a  38 6-hour-older GFS forecast. Second, the forecast skill of a model is a function of wavelength. It decreases faster for shorter waves; most of the longer-range skill is in the larger waves. In a mesoscale model with high resolution topography, for instance, you can get good details on the rainfall distribution when a moist system moves across the mountains. But if the system is poorly forecast, then those details will all be meaningless for a daily forecast. In other words, if the big picture is wrong, quickly for a mesoscale model and gradually approaches that of a coarser resolution model. Links Characteristics of Operational NWP Models (COMET) http://www.meted.ucar.edu/nwp/pcu2/launpcu2.htm