a. NGM and Eta Model forecast assessments

Figure 1b shows the observed low at 1800 UTC 14 September 1996, when the system was near its peak intensity. It is appropriate to assess the model runs that were initialized 30 h earlier, at 1200 UTC 13 September 1996, because participants in the sailboat race that began the next morning were still deciding early that Friday evening whether or not to participate. Additionally, these model runs were consistent with the MM5 runs that will be described in the next section that had to be initialized at this time. A later initialization time would have been within the period of maximum deepening and hence would have resulted in the no-lake run being contaminated by lake-aggregate heating that would have already modified the atmosphere.

A careful surface analysis at 1200 UTC 13 September 1996 (e.g., Fig. 5a in HH) revealed a complicated situation with an occluded front separating two 1004 hPa lows: one over Lake Huron and another one over eastern Pennsylvania. A third weak low was located just off the southern New Jersey coast. Figure 2a shows that the NGM initialization included all three lows: one low over Lake Huron with a central SLP of 1004 hPa, one low in Virginia with a central SLP of 1004 hPa, and one low farther offshore. During the first 12 h of the NGM forecast, the low over Lake Huron weakened as the one over Virginia retrograded northwestward to a position north of Lake Ontario. As a result, the NGM forecasted position of the primary surface low was close to the observed position at 12 h (0000 UTC 14 September 1996) over Lake Huron, but then the NGM low proceeded to slow its westward progression relative to the observed low over the next 30 h. The forecasted center remained east of Lake Huron throughout the 48-h period, which is significant given the storm’s observed rapid intensification once it moved over water. For example, Fig. 2b shows that the 30-h NGM forecast, valid 1800 UTC 14 September 1996, had the center over Georgian Bay, at the base of the Bruce peninsula, while the observed center was located over southern Lake Huron (Fig. 1b). Additionally, the central pressure was about 6 hPa too high (999 vs 993 hPa) during the most rapid intensification and peak intensity of the storm (1800 UTC 14 September 1996). Weakening also took longer in the NGM. By 1200 UTC 15 September 1996, the observed central pressure had risen to 1005 hPa, while the NGM kept the central pressure at 999 hPa (not shown).

Figure 3a shows the Eta Model initialization, which included a 1004-hPa surface low just east of Buffalo and another one just off the southern New Jersey coast. A weak trough was located over Virginia. During the first 24 h of the Eta forecast, the two lows merged over central Pennsylvania and then slowly retrograded northwestward as a single low to a position between Lakes Huron and Ontario. This position was as far west as the Eta forecasted the low to move, this is farther east than was observed and even farther east than it was in the NGM (Fig. 3b). During the 24–48-h forecast period, the low was forecasted to move eastward too quickly. Despite the position errors, the forecast intensity of the surface low was closer to the observed than that from the NGM. At 1800 UTC 14 September 1996, when the observed low was most intense, the Eta Model low was 300 km too far east and the central pressure was 2 hPa too deep (e.g., 991 hPa). The fact that the low was too deep was a characteristic that developed immediately, once the two lows in the initialization merged over central Pennsylvania. The performance of the two models in terms of the observed low over Lake Huron may be summarized as follows: the NGM low position was 200 km too far east and the intensity was 6 hPa too weak; the Eta low position was 300 km too far east and the intensity was 2 hPa too strong.

The NGM and Eta Model forecast errors in the position and intensity of the surface low contributed to (near) surface wind direction and speed errors. Wind speed errors over land were generally less than 3 m s⁻¹ in both models but more significant over water. For example, at 1800 UTC 14 September 1996, winds predicted by the NGM were only about 5–10 m s⁻¹ over Lakes Huron and Erie. Winds predicted by the Eta Model were 8–10 m s⁻¹ over Lake Huron and 10–13 m s⁻¹ over Lake Erie. Observed winds were 15–18 m s⁻¹ with gusts to 23 m s⁻¹ over both lakes (Fig. 1b and Fig. 7b in HH).

An analysis of observed (station measured not grid averaged) 48-h cumulative rainfall ending 1200 UTC 15 September 1996 indicated that both the NGM and Eta forecasts underestimated precipitation totals near Lakes Huron and Erie (Fig. 4). For example, reports of more than 5 cm (2 in.) of rain in the thumb of Michigan (e.g., along the southwestern shores of Lake Huron) and in western New York were common with some isolated reports of 10 cm (4 in.). In contrast, the NGM forecasted amounts for the same time period in these regions were about 2.5 cm (1 in.). The Eta forecasted amounts were better, having forecasted maximum total precipitation amounts over 5 cm over Lake Ontario and southern Ontario. The Eta Model did seem to simulate to some degree the precipitation enhancement along the downwind lakeshores, but the heaviest amounts were too far east, along the north shore of Lake Ontario.

A likely cause for the NGM and Eta Model errors is the coarse grid spacing, which was insufficient to resolve the meso-β-scale precipitation signatures in the observed precipitation distribution along the lakeshores. The spacing may have been too coarse to simulate properly the convergence/divergence patterns over the individual lakes that resulted from the surface fluxes of heat and moisture if not the surface fluxes themselves. The 80-km grid spacing in the NGM meant that only two to three grid points were located over Lake Huron in the east–west direction. This resolution is insufficient to simulate properly the heating (effects) from the individual lakes, as was evidenced by individual lake-scale troughs that were barely discernable over the lakes during the forecast period. The 48-km grid spacing in the Eta Model meant that three to four grid points were located over Lake Huron. This resolution is slightly better than that of the NGM at resolving the heating (effects) from the individual lakes, so effects in terms of generating individual lake-scale troughs were only slightly more evident.

Another potential source for errors is the parameterization schemes. Gallus (1999) evaluated the performance of the Eta Model using different resolutions and different convective schemes. He found that one feature of the Betts–Miller–Janjic convective scheme in that model is that it can be prematurely triggered, which can result in too much (convective) precipitation, too much gridscale drying, and too little subsequent gridscale precipitation. The release of too much latent heat early on may have been the case with the Eta Model. This hypothesis is supported by the facts that the Eta low developed more rapidly than observed, after it merged with another low over central Pennsylvania, and that the model generated too much convective precipitation along the north shore of Lake Ontario. The premature (convective) development may have been fueled by strong surface fluxes from Lakes Erie and Ontario, the subsequent convective release of latent heat, and the associated sea level pressure falls, which may have prevented the low from developing farther west, over Lake Huron.

Other possible causes for errors in both models may have been strong horizontal diffusion that can force synoptic-looking precipitation distributions, improper surface low initialization, incorrectly initialized LSTs, or failure to account for the observed changes in LSTs during the period. Buoy 45008 in southern Lake Huron reported nearly a 7°C decrease over a 36-h period of the event. About half of that occurred in one 4-h period (see HH for details). The impact of this temperature decrease on surface fluxes and the evolution of the surface low will be described in more detail in section 3. The interested reader is referred to Hoke et al. (1989) and Black (1994) for more complete descriptions of the NGM and Eta Models, respectively.

b. Textual forecast assessment

The textual forecasts issued by the National Weather Service Forecast Office (NWSFO) in White Lake, Michigan (DTX), and the Canadian Meteorological Centre in Toronto, Ontario, closely paralleled the operational model forecasts. The forecasts from these two offices were similar, so only the ones from DTX are described.

The marine forecast that was issued from DTX for Lake Huron and Lake St. Claire was certainly an improvement over that from the raw model output, but it still underestimated the observed wind speeds. For example, a small craft advisory was issued on Friday afternoon (13 September 1996) for the coastal waters of Lake Huron. Northwest winds between 8 and 13 m s⁻¹ and wave heights between 1 and 2 m were expected throughout the ensuing 24–36-h forecast period. Recall that observed wind speeds around the center of the storm were near 15 m s⁻¹ with gusts to 23 m s⁻¹ and wave heights were near 3 m as reported by the two buoys on Lake Huron (Fig. 7b in HH). Numerous sailboats that were participating in the Bayview “Night Race,” which is a 50 nautical mile race on Lake St. Claire, also reported sustained winds throughout the afternoon between 15 and 18 m s⁻¹ with gusts to 23 m s⁻¹. Several boats were demasted, and at least one person suffered a concussion as a result of the high winds and very choppy lake surface. Gale warnings and increased wave height forecasts were not issued by DTX until 2100 UTC 14 September 1996, when the system had already passed its peak intensity. In contrast, the NWSFO at Buffalo, New York, had called for these strong winds and waves earlier in the period for eastern Lake Erie, starting with Friday morning’s forecasts, primarily due to an intense lake-effect situation and intense vertical mixing that was anticipated over that region. However, the strong winds that were forecasted for that region did not occur because the system was farther west than anticipated.

In terms of precipitation, the forecast issued by DTX for the Great Lakes region called for mostly showery precipitation with some areas of steadier rain closer to the low center. Recall that many places in the region received 0.635–1.25 cm (0.25–0.5 in.) of rain, except for portions of the thumb and Ontario along the southeastern Lake Huron shoreline, where heavier amounts occurred. Localized flooding occurred in these areas as more than 5 cm (2 in.) of rain fell. Flood warnings were later issued for this area upon receiving these reports (advisories not shown). In contrast, the NWSFO in Buffalo, New York, had called for frequent and significant lake-enhanced rain over western New York where it was thought a significant early season lake-effect event would take place. Rain amounts in excess of 5 cm (2 in.) were reported by spotters, which prompted flood warnings to be issued for areas surrounding Lake Erie.

a. Model information and initialization

A fine grid mesh (FGM) with 20-km grid spacing was surrounded by a coarse grid mesh (CGM) with 60-km grid spacing. The grids were run interactively; information from the FGM was fed to the CGM and the CGM provided boundary conditions for the FGM. The 20-km spacing in the FGM provided 9–10 grid points across Lakes Superior and Huron from north to south and six to seven grid points across Lake Michigan from east to west. The model used 35 vertical (sigma) levels with a model top pressure of p_t = 100 hPa. Most of the levels were chosen to be below 700 hPa in order to capture important boundary layer processes. The Blackadar (Blackadar 1979) high-resolution planetary boundary layer scheme was used to simulate conditions near the surface, the Kuo scheme was used to model convection in the CGM, and the Kain–Fritsch (Kain and Fritsch 1990, 1993) scheme was used to model convection in the FGM. Sousounis and Fritsch (1994) used the Kain–Fritsch scheme to simulate a cold air outbreak in November 1982. The choices for the two different convective schemes for the CGM and FGM were appropriate for the resolutions of these domains (Kain and Fritsch 1998).

The model was initialized using surface and upper air data on mandatory levels valid at 1200 UTC 13 September 1996 from the National Centers for Environmental Prediction (NCEP) Historical Archives. The initialization procedure was similar to that described in Sousounis and Fritsch (1994) for their MM4 simulations of a cold air outbreak case in November 1982 (hereafter NOV82). Lake surface temperatures were initialized using satellite-derived data from the Great Lakes Environmental Research Laboratory and buoy data from the National Data Buoy Center. No special or subjective preprocessing (e.g., bogusing), data assimilation, or changes in parameterizations were implemented for this MM5 simulation. More information about the MM5 model is available in Dudhia (1993) and Grell et al. (1994).

b. Model simulation and verification

Figure 5a shows that the MM5 FGM initialization of the SLP field at 1200 UTC 13 September 1996 revealed two 1005-hPa low centers: one over southern Lake Huron and one over northwestern Pennsylvania. The CGM initialization additionally revealed two troughs outside the FGM: one extending off the coast of southern New Jersey and the other extending into Virginia. The MM5 initialization was very close to those in the NGM and Eta Model. The largest difference may have been the weak center of low pressure in northwestern Pennsylvania, which did, however, exist as a weak center of low pressure in south-central Pennsylvania in the surface observations (see Fig. 5 in HH). Other features in the MM5 initialization such as the troughs over Lake Superior and central Ohio also compared well with those in the observations. Their existence in the MM5 initialization was probably a result of the finer grid spacing than that which existed in the other two models. By 12 h (0000 UTC 14 September 1996), the MM5 SLP field showed a more coherent (single) low structure north of Lake Ontario, a behavior similar to that found in the operational models. Over the next 18 h, the MM5 forecasted low retrograded slowly, to a position near Lake Huron. By 1800 UTC 14 September 1996, the low was located along the eastern shore of Lake Huron (Fig. 5b), about 70 km to the east of the observed low for that time (see Fig. 1b). The MM5 SLP minimum was within 1.7 hPa of the observed minimum. During the remainder of the 48-h period, the MM5 low moved more slowly eastward and weakened more slowly than the observed one.

The MM5 forecasted winds in the lowest sigma layer of the model, which was about 20 m above ground level, show close agreement with observed surface winds. Specifically, winds over both land and water were generally within 2–3 m s⁻¹ of the observed values. Winds over land were weaker because of increased friction and reduced surface heating. Winds over the lakes were considerably stronger because of reduced friction and increased surface heating. It is particularly noteworthy that the MM5 forecasted accurately the observed wind speed maximum of 15 m s⁻¹ over southern Lake Huron near buoy 45008 during the time of peak intensity at 1800 UTC 14 September 1996 (Fig. 5b). It is important to mention that the buoy winds on the Great Lakes are measured at 5 m above the lake surface and that observed winds at a height of 10–20 m could have easily been 10%–20% higher. That is, a 16 m s⁻¹ wind at 5 m could have been a 19 m s⁻¹ wind at 20 m, and could have accounted for some of the differences between the MM5 forecasted and observed wind speeds over the lakes. Recall that some of the sailboats on Lake St. Claire, just to the south, had reported winds near 18 m s⁻¹ around that time. These winds were measured for the most part at mast height between 10 and 30 m.

Recall that both the NGM and Eta Model forecasted considerably weaker winds over southern Lake Huron. As a result of good agreement with observations regarding not only surface winds but also surface temperatures and moisture, the surface fluxes that were generated by MM5 compared well with those determined from observations. Table 1 shows surface fluxes from observational data at buoy 45008 and from NGM, Eta Model, and MM5 output taken at the corresponding grid point. Because no moisture information was available from the buoy, a climatological value for relative humidity of 80% (Saulesleja 1986) was assumed at anemometer level. Also, for consistency, the model fluxes in Table 1 were determined using bulk aerodynamic formulas. Two features are important to mention. First, the observed fluxes at the beginning of the time period were significantly higher than those from any of the models. This aspect was likely the result of a cold front that had just passed the buoy (45008) so that winds had a significant ageostrophic component associated with them (directed across the front) and hence a higher speed that was not captured in the model initializations. Second, all of the models generated fluxes that were significantly higher than those observed at the end of the period. The higher fluxes were probably a result of the fact that none of the models accounted for changes in LSTs but could also be simply a result of the missed location of the low center with respect to the buoy. Because lake–air temperature differences during the period of strongest development were 10°–12°C, the observed 5°–6°C drop in lake surface temperatures at buoy 45008 alone could have accounted for a 50% reduction in combined sensible and latent heat fluxes (∼150 W m⁻²) at the surface (see Fig. 6 in HH). The fact that the surface air became less unstable as a result of the reduced lake–air temperature difference and that stronger winds could not mix down to the surface could account for the 50% reduction in wind speeds and an additional 50% reduction in fluxes. It is also likely that, as the surface winds and fluxes weakened, the central pressure rose and weakened the pressure gradient, further weakening the winds and fluxes. These likelihoods are supported by the fact that surface winds did decrease abruptly over buoy 45008. These two effects likely reduced the observed fluxes significantly at the end of the period.

An analysis of the precipitation forecast shown in Fig. 6 indicates that the MM5 was superior to both the Eta Model and the NGM. For example, the MM5 generated storm total precipitation amounts between 0.635 and 2.5 cm (0.25–1.0 in.) over east-central lower Michigan, and amounts over 10 cm (4 in.) along the southeast shore of Lake Huron in Ontario and northeast of Lake Erie. Model signatures of lake-effect rainbands were also produced along the shores of Lakes Michigan and Erie, in which the Lake Erie band is confirmed by the greater than 13 cm (5 in.) of rain reported in Buffalo, New York (Fig. 4). Lighter amounts were observed on the southeast shores of Lake Michigan.

The MM5 forecast of the position and intensity of the 500-hPa low at 48 h (1200 UTC 15 September 1996) was within 100 km and 1 dam, respectively, of the NGM analyzed (observed) 500-hPa low (Fig. 7).¹ The NGM forecasted low was approximately 200 km farther to the northeast and 2 dam deeper than observed. The Eta forecasted low was better located but 4 dam deeper than observed. The relative position and intensity errors in the NGM and Eta Model mirror those that existed at the surface. The higher, more accurately forecasted, 500-hPa heights by the MM5 were likely the result of the heat and moisture from the lake aggregate being distributed correctly (and more deeply) through the atmosphere. Other features at 500 hPa demonstrate the greater accuracy of the MM5 forecast. For example, the northwestward extension of the 500-hPa trough across the western lakes region that existed in the analysis was captured for the most part by MM5. The NGM had only a weak version of this extension, while the Eta Model did not show it at all. Additionally, the 10 m s⁻¹ northwesterly winds over southern Lake Huron were also captured well by MM5. Both the NGM and the Eta Model had forecasted winds to be nearly 15 m s⁻¹ at this time. The weaker winds in the analysis and MM5 forecast may have been the result of deep convective mixing, which may have resulted because of the relatively warm lake surface. This deep mixing not only distributed the heat over a great depth and caused a weaker height gradient and hence weaker geostrophic winds, but it also likely reduced the vertical wind shear. Some evidence that the MM5 simulated the deep mixing better than the other models can be seen by comparing the temperature distributions from these models with the observed one. For example, the NGM had −20°C air surrounding the 500-hPa low and the Eta had −16°C air surrounding the 500-hPa low, while the MM5 and the analysis both had −18°C air surrounding the 500-hPa low at 48 h. In fact, the analysis and the MM5 temperature distributions show a very weak thermal gradient surrounding the low center. While the suggestion of a warm core is evident, it is not visible from the contours. A warm core is more evident in the MM5 temperature distributions at lower levels (e.g., 700 hPa). In contrast, the other two models have very different temperature distributions. The NGM model appears to have a cold core surrounding the low center and the Eta Model appears to have a very baroclinic structure with cold and warm advection on either side of the low center.

Other aspects of the vertical structure of the low were also forecasted well, even at 48 h, despite the higher than observed surface fluxes at buoy 45008. Figure 8 shows that the 48-h model forecast sounding for Buffalo, New York, which is valid at 1200 UTC 15 September 1996, compares very well to the observed sounding. Small differences exist; for example, the intricate structure near the top of the planetary boundary layer (750–700 hPa) is not precisely replicated in the model. However, the MM5 temperatures are within 1°C of the observed temperatures everywhere else below 400 hPa. At 300 hPa, the observed sounding is 2°C cooler than the modeled sounding, which is likely a result of the MM5 inversion being 25 hPa lower than observed. The lower than observed model tropopause suggests that the lake-aggregate heating and moistening was not extending as high into the atmosphere as observed. The accuracy of the temperature structure over Buffalo at 48 h, as well as many other aspects of the forecast, are especially important to demonstrate the efficacy of this model (forecast) because it is over this (eastern Great Lakes) region where the lake-aggregate heat and moisture was accumulating and hence where the impacts of the lakes were likely greatest at the end of the forecast period.

The MM5 forecast more closely resembled the observations than did either the Eta Model or the NGM in many variables, perhaps because of its finer grid spacing. In order to demonstrate further the importance of resolution, an additional MM5 simulation was performed using just one CGM domain with 60-km grid spacing. This simulation showed a weaker surface low, with a minimum SLP of 1001 hPa, within 50 km of the one in the two-way interactive FGM version. Maximum surface wind speeds over Lake Huron were 5–10 m s⁻¹. The results from this stand-alone CGM simulation suggest that the higher resolution in the coupled MM5 run was necessary for capturing the minimum in the central mean sea level pressure and the higher surface winds that were observed, but not necessarily the location. The fact that the stand-alone CGM MM5 simulation forecasted the location of the low very accurately adds credibility to the hypothesis that the errors in the Eta Model forecast may have been compounded by other features in the model besides the 48-km grid spacing.

Warner and Seaman (1990) and Ballentine et al. (1998) have demonstrated the improvement that a high-resolution mesoscale model can provide during lake-effect snow situations over the lakes. It is likely that high resolution was beneficial in this case as well because surface fluxes over the lakes were important. This benefit may have stemmed from two sources. First, while higher resolution by itself does not necessarily lead to a better gridpoint representation of surface fluxes of heat, moisture, and momentum, it does provide for more water points and it does likely create the potential for the model to generate more realistic boundary layer heating, more realistic convergence zones, more realistic vertical motion, and more realistic precipitation in situations when cold air flows across warm lakes. Second, this more realistic representation of individual lake effects means a better representation of aggregate effects, which may include the development or enhancement of a meso-α-scale cyclonic circulation at the surface, and associated meso-α-scale convergence, vertical motion, and precipitation. The meso-α-scale circulation may then contribute to the strength of the fluxes over the individual lakes in a positive feedback loop as mentioned earlier. In short, a high-resolution numerical weather prediction model may perform better over the lakes in unstable situations because it simulates more accurately the multiscale impacts of the Great Lakes.

It is certainly a valuable exercise to explore even further and understand more conceptually the causes for the better performance by MM5 by performing a series of MM5 and/or Eta Model simulations with various resolutions and parameterizations. However, because that exercise is not really within the scope of this paper, the next section will examine some specific aggregate impacts of the lakes and describe possible mechanisms that influenced the characteristics of the low.

a. WL–NL sea level pressure and surface wind differences

Figure 9 shows the intensities and positions of the WL, NL, and observed surface lows at selected times. By 1800 UTC 14 September 1996 (30 h), the WL low was 5 hPa deeper than the NL low. Additionally, the WL low executed a tighter trajectory, remained to the east of, and moved more slowly than the NL low. Casual inspection of Fig. 9b may suggest that the NL simulation was a better match to the observed conditions than the WL simulation because the NL low was closer to the observed one than was the WL low at 30 h. This feature, however, does not invalidate the simulations. Rather, it suggests that a combination of strong surface fluxes, frictional convergence, rising motion, and latent heat release in the WL simulation may have caused the low to remain closer to the lakeshore for a period of about 6 h instead of moving farther westward over the open water before (stronger) dynamical processes began to effect an eastward movement. Inadequately specified LST gradients may have further contributed to WL track errors. Importantly, the tighter trajectory in the WL simulation suggests that the aggregate heating (and frictional convergence at the surface near the Lake Huron shoreline) provided a greater focusing mechanism for the low, while in the NL simulation the low was steered more by dynamic processes aloft such as vorticity advection from a short wave at 500 hPa (see Fig. 5 in HH). More careful inspection of Fig. 9a reveals the fact that the central SLP of the WL low was less than 2 hPa higher than that in the observed (995.2 vs 993.5 hPa), while the central SLP of the NL low was nearly 7 hPa higher (for the same times). Furthermore, considering the fact that the NL simulation was initialized at 1200 UTC on 13 September, when cold air was already overspreading the lakes, it is likely that the NL low had some residual lake-induced heating at the start of the simulation, so that the WL–NL differences (for many fields) were likely underrepresented. The initialization time of 1200 UTC 13 September 1996 in fact was considered to be a good compromise between starting the simulation at an earlier time, which would have eliminated heat contamination but increased the likelihood of other sources of error, and starting the simulation at a later time, which would have exacerbated the heat contamination problem. The weaker surface low and the absence of (smooth warm) water in the NL simulation were the primary reasons for the lower NL wind speeds, which were nearly 10 m s⁻¹ less than the WL wind speeds over Lake Huron.

Strong easterly flow during the last 12 h, from 0000 to 1200 UTC 15 September 1996, likely resulted in upwelling across eastern Lake Huron and reduced LSTs (see Fig. 8c in HH). These lower LSTs led to reduced surface fluxes and hence to a reduced tendency for the lakes to constrain the movement of the low, which by this time was eastward once again as indicated by observations and by the NL simulation. Because the MM5 model currently does not account for changes in LSTs during a forecast/simulation, the surface fluxes were overforecast during this period, which likely resulted in the low continuing to remain close to the lakes rather than to begin moving eastward.

b. Vertical structure of the WL–NL perturbation

An examination of, and even an explanation for, the WL–NL SLP differences provides only a partial description of how the lake aggregate affected the synoptic-scale low. More insight to the impacts of the lakes on this system can be obtained by examining WL–NL differences (e.g., perturbations) at selected upper levels. For example, at 850 hPa, the heat and moisture from the lake aggregate spread upward and outward quickly. By 24 h, perturbations in equivalent potential temperature (θ_e) and wind extended across the Great Lakes. By 48 h, θ_e perturbations exceeding 2 K covered most of the Great Lakes region; strongest θ_e perturbations ranged from 8 to 10 K over certain locations of western Ontario (see shading in Fig. 10a). A closed cyclonic perturbation circulation, with a width near 700 km, was centered over eastern Lake Huron where the strongest θ_e perturbations existed and demonstrated the lake-aggregate-scale nature of the perturbation. Perturbation winds just north and south of the center exceeded 10 m s⁻¹.

The perturbation took a little longer to reach 500 hPa, but by 24 h, θ_e perturbations of 2 K covered small areas near Lakes Michigan, Huron, Erie, and Ontario. In the ensuing 24 h, however, the perturbation spread quickly over a large region. By 48 h, θ_e perturbations exceeding 1 K covered most of the eastern Great Lakes, extending southeastward from eastern Lake Superior to northwestern Pennsylvania. Strongest θ_e perturbations ranged from 4 to 6 K over Lake Ontario (see shading in Fig. 10b). The region of positive θ_e perturbation coincided with a well-defined anticyclonic perturbation circulation with a width near 700 km. Perturbation wind speeds exceeded 10 m s⁻¹ near the center.

Perturbations did not take much longer to reach 300 than 500 hPa. By 24 h, negative θ_e perturbations exceeding −1 K spanned a region that covered southern Lake Huron, Lake Erie, Lake Ontario, and western Ontario. The negative θ_e perturbations that existed over eastern Lake Huron were stronger than the positive ones that existed at 500 hPa for the same time. By 48 h, negative θ_e perturbations exceeding 1 K extended southeastward from a region just north of Lake Superior to western New York and negative θ_e perturbations from −4 to −7 K existed over southeastern Ontario (see shading in Fig. 10c). The region of negative θ_e perturbations coincided with a well-defined anticyclonic perturbation circulation with a width near 700 km. Perturbation wind speeds near 8 m s⁻¹ existed near the center, which was located just north of Lake Ontario.

The WL–NL perturbations at 48 h were primarily the result of magnitude (e.g., Lagrangian) differences. For example, recall from Fig. 9a that the magnitude difference at the surface was approximately 5 hPa (or equivalently, 40 m at 1000 hPa) that had developed by 24 h. The magnitude differences that developed at upper levels were slightly less and occurred later. For example, Fig. 11 shows WL and NL 300-hPa heights, winds, and temperatures valid at 27 h (1500 UTC 14 September 1996), just before the surface low reached peak intensity. A comparison between the two illustrates how the WL heights were only just beginning to differ from the NL ones at 300 hPa. Specifically, heights were slightly (e.g., 20 m) higher, more (geostrophic) diffluence was clearly evident over southern Lake Huron, weaker winds covered that region, a warm core was developing over southern Lake Huron, and a ridge, extending from southeastern lower Michigan to the center of the 300-hPa low, was beginning to appear in the WL simulation. The higher heights and the ridge in the WL simulation at 300 hPa shown in Fig. 11 (and at 500 hPa; not shown) were likely the result of deep diabatic heating from the lakes, similar to what happens with a hurricane or any warm core vortex where a deep warm layer near the center of the circulation causes an upward deformation of height surfaces above the level of maximum warm perturbation or heating and anticyclonic wind shear results in reduced cyclonic or (increased) anticyclonic winds at upper levels.

The developing magnitude differences also contributed to location differences. By 48 h, the 300-hPa WL low was centered over eastern Lake Huron with a height of 916 dm while the 300-hPa NL low was centered only 150 km to the east-southeast, over Lake Ontario, with a height of 911 dm. In fact, the NL 300-hPa heights fell from 913 to 911 dm while the WL 300-hPa heights rose from 913 to 916 dm over the 48-h period. The higher heights (and deeper vertical mixing) in the WL simulation also meant that slightly weaker winds, vorticity, and vorticity and temperature advections existed at mid- and upper levels. It is likely that the weaker dynamic forcing at upper levels and the continuous thermal forcing from the lakes over a fixed region in the WL simulation combined to cause a slower and more confined movement of the WL low at all levels at the surface for most of the 48-h period.

The heating and moistening from the Great Lakes was so extensive that it did not just warm the center of an extratropical system, but rather converted a cold core system into a warm core system. Figure 12 is a CGM vertical cross section of WL and NL temperature structures across a region extending eastward from North Dakota to a point located several hundred kilometers east of Cape Cod, Massachusetts, valid at 48 h. The deep synoptic-scale cold core (outer shaded region) of the air mass is clearly evident in the NL simulation. Going from west to east along constant-pressure surfaces across the region, temperatures drop abruptly over Minnesota and then rise abruptly over Massachusetts. The WL temperature structure, however, reveals a warm core (embedded within the cold core)—a sharp temperature rise over the western edge of the lakes, a peak over western New York, and then a gradual decline toward the east—especially at low levels. The width of the warm core (inner shaded region) at any given pressure level is determined by the locations where the horizontal WL temperature gradient is zero. The strength (darkness of inner shading) is determined by the temperature difference between the maximum WL temperature (in the middle) and the average WL temperature at the edges. The warm core weakens near 600 hPa, becomes stronger at 450 hPa, and then weakens again around 350 hPa. Nowhere is the core much stronger than ∼2°C. The weakening around 600 hPa is likely a result of the warming being spread over a larger area so the horizontal temperature distribution is flat rather than exhibiting a localized maximum. The cooler conditions in the WL simulation above 350 hPa are likely the result of 1) the tropopause having been lifted and stable air having been adiabatically cooled, 2) mesoscale upward motion, or 3) convective overshooting.

c. Differences between this case and a wintertime case

A cold air outbreak that occurred over the Great Lakes during 13–15 November 1982 (NOV82) has been studied extensively (Sousounis and Fritsch 1994; Sousounis 1997, 1998). It provides an opportunity, through comparison with the present case, to understand more about the sensitivity of aggregate effects to synoptic conditions. The NOV82 case began with a strongly baroclinic low moving northeastward through the Great Lakes region, bringing cold air southward across the lakes. The cold air persisted, even as a synoptic-scale high moved east across the region and the surface flow changed from northwesterly to southwesterly to southeasterly ahead of an approaching trough. The flow then became northwesterly once again as the weak trough developed into a closed low and moved northeastward across Lake Huron. Combined surface sensible and latent heat fluxes were 200–500 W m⁻² for much of the 48-h period (see Fig. 3 in Sousounis 1998). Lake–850-hPa air temperature differences were −20°C (e.g., lake surface temperature for Lake Michigan was +7°C and 850-hPa temperature was −13°C at Green Bay, Wisconsin, at 1200 UTC 13 November; see Fig. 4a in Sousounis and Fritsch 1994). Snow showers were widespread across the region for much of the time (Sousounis and Mann 2000). The NOV82 case was one studied by Weiss and Sousounis (1999) in their climatological evaluation of wintertime effects of the Great Lakes aggregate. The WL–NL SLP perturbation for the NOV82 case was around the median value of 6 hPa, which they found from their study that included more than 200 cases over a 10-year period.

A comparison of the WL–NL perturbations for the SEP96 and NOV82 cases shows distinct differences. First, the WL–NL potential temperature and wind perturbations were significantly stronger in the SEP96 case than those in the NOV82 case. Perturbation θ_e values in the current case were as high as 12°C at 850 hPa at 36 h. Perturbation θ_e values in the NOV82 case were only 6°–7°C at 850 hPa. Second, WL–NL perturbations extended to depths that were considerably greater in the SEP96 case than those in the NOV82 case. Perturbation θ_e values exceeding −2°C in magnitude and perturbation wind speeds exceeding 5 m s⁻¹ in the current case extended to 200 hPa. Such perturbations were basically nonexistent above 600 hPa in the NOV82 case. Third, the anticyclonic perturbation flow collocated with positive isentropic differences at 500 hPa were opposite to those in the NOV82 case at any level, where either cyclonic perturbation flow typically accompanies positive isentropic perturbations (near the surface) or anticyclonic perturbation flow typically accompanies negative isentropic perturbations (near 700 hPa).

The unique features in the present case are illustrated from a different perspective in Fig. 13, which shows at 48 h a vertical cross section of the FGM WL–NL height and θ_e perturbations extending from Madison, Wisconsin, to Augusta, Maine. Three aspects are noteworthy because they are distinctly different from those in the NOV82 case, a cross section of which is included in Fig. 13 for comparison. First, the fact that negative height perturbations extended above 700 hPa and that positive height perturbations extended as high as 200 hPa in the current case indicates that not only was the WL–NL response deeper than that in the NOV82 case, but that the level of maximum lake-aggregate heating was above 700 hPa. Second, the amplitude of the positive height perturbations at 400 hPa was greater than the amplitude of the negative height perturbations at the surface. Again, this feature is different than what was found in the NOV82 case, which may be related to the lower stability and to the fact that height displacements are easier to achieve at higher altitudes than at lower altitudes because of lower pressure. Third, the fact that a warm core in the perturbation θ_e field extended to 400 hPa indicates the high altitude to which the perturbation extended. Note in the NOV82 case how the warm plume only extended to 700 hPa and how it was tilted severely downwind because of the strong winds above the surface that were associated with low-level baroclinicity and the eastward advancement of an open wave.

The slow movement of the system, the strong surface heating, and the low static stability likely contributed to the unique WL–NL differences in this case. Additionally, surface frictional differences between the WL and NL simulations likely contributed to differences in characteristics of the WL and NL lows directly above the lakes themselves. The slow movement underscores two important differences between this system and typical wintertime systems (or any other systems for that matter) that move through the Great Lakes Region. First, the generally slow movement of this system was related to a large degree to the cut-off nature of the system and to a lesser degree to the influence of the lakes. Wintertime systems are typically associated with more (background) baroclinicity, have a more open wave structure, and hence move more rapidly across the region. The fact that the residence time over the lakes can be considerably greater for cut-off (nearly barotropic) systems than for open wave (baroclinic) systems means that the time-integrated lake-aggregate effect can be significant for cut-off lows, even if the other conditions such as lake–air temperature difference are not as favorable. Second, because this system was associated with very little wind shear, the heat and moisture from the lakes remained concentrated rather than spread over a large region, as is commonly the case with typical wintertime systems.

Strong surface heating also contributed to the WL–NL differences. For example, the WL simulation exhibited LSTs as high as ∼20°C, surface air temperatures as low as 10°C, and strong (gusty) surface winds exceeding 17 m s⁻¹ over southern Lake Michigan and southern Lake Huron. The NL simulation exhibited ground and surface air temperatures that were nearly identical and surface winds around 7 m s⁻¹ over the same region. The WL–NL differences for the combined surface sensible and latent heat fluxes approached 700 W m⁻² at times for the region. While heat fluxes of this magnitude are typical for winter during cold air outbreaks across the Great Lakes, the ratio of latent to sensible heat flux is much different. Figure 14 shows that latent heat flux differences exceeded sensible heat flux differences in the SEP96 case by a factor of 2 at 24 h over all of the lakes except for Lake Superior. The Bowen ratio B ∼0.5 was typical throughout much of the 48-h period. Bowen ratios B ∼ 0.1 are characteristic of tropical cyclones (Anthes 1982). In contrast, during typical winter conditions when the air and lakes are colder, much lower saturation vapor pressures contribute to much lower latent heat fluxes so that the sensible and latent heat fluxes are more nearly equal (B ∼ 1). Such high Bowen ratios are characteristic of polar lows (Shapiro et al. 1987). Because latent heat is released upon condensation (in clouds), the deeper clouds and higher saturation vapor pressures in the SEP96 case contributed to a deeper lake-aggregate response than for typical winter cases. Additionally, the deeper heating and higher level of maximum heating likely contributed more effectively to SLP falls (Hirschberg and Fritsch 1993).

Low static stability in the current WL simulation also likely contributed to the strong surface heat fluxes warming a deep layer of the atmosphere. A comparison of WL and NL soundings for a representative time and location² over Lake Huron (Fig. 15) illustrates that the temperatures in the WL sounding were higher than those in the NL sounding by several degrees from the surface to ∼400 hPa. The fact that the (warmer) WL temperature sounding nearly follows a moist adiabat over a considerable depth suggests that deep convection was occurring over the region. The WL–NL temperature differences reverse above 400 hPa because the WL tropopause is ∼100 hPa higher than the NL one. Importantly, the NL sounding is less stable relative to that for typical wintertime cases (cf. Figs. 15a and 15b). The low stability and the strong surface fluxes explain the great depth of the WL–NL perturbations. Table 2 summarizes some of the synoptic-scale environmental differences as well as the perturbation differences between this case and the NOV82 case. The environment values are obtained from NL output averaged over the center of the perturbation and the perturbation values are obtained from WL–NL differences near the time of maximum WL–NL surface differences.