On 8–9 October 2017, more than 170 wildfires ignited in the Wine Country, northern coastal ranges, and Butte and Nevada Counties to the west, north, and east of California’s northern Sacramento Valley (Figure 1
). Of these, fourteen large fires grew rapidly, some joining into multi-fire complexes. Subsequent investigative reports [1
] determined several of the large fires were started by electrical equipment, in some cases by branches being brought into contact with those. Many fires appeared and rapidly spread during local peaks of an unusually strong downslope wind event. Such multi-day events, termed Diablo winds, are a recognized meteorological feature of the region [3
] and have been linked to erratic wildfire behavior, fatalities, and past destructive wildland–urban interface fires such as the 1991 Oakland Hills Fire [5
] and the similar 1964 Wine Country fires [6
]. In 2017’s event, the widespread outbreak of fires and their extremely rapid spread raise questions about the airflow regime that occurred and the underlying mechanisms leading to the extreme winds. The events hint at exceptional wind extrema, yet these lie between surface weather station network data locations. While regional elevated winds are captured in operational simulations, if there are extrema, they could be under-resolved in operational forecasts. We examined these issues using two approaches. First, we performed a retrospective research simulation with the CAWFE®
(derived from Coupled Atmosphere–Wildland Fire Environment) coupled numerical weather prediction–wildland fire behavior modeling system. Second, we configured CAWFE without a priori knowledge of how the fire unfolded in order to predict the flow regime and subsequent Tubbs Fire growth, examining the predictability of the winds including their extrema and fire behavior.
2. Overview of Event Environment
Diablo winds are a San Francisco Bay meteorological phenomenon that arises from concurrent strong high pressure over the Great Basin and lower pressure offshore of San Francisco and Monterey. Similar to Santa Anas, this inland–offshore pressure gradient drives air downward in elevation and offshore. Climatologies developed from surface weather station data suggest that Diablo winds tend to occur overnight through morning in fall through spring [7
]. They are characterized by low relative humidity and high wind speeds, but temperatures and ratio of wind gusts to mean wind speeds are similar to climatology [7
], presenting a dry, potentially strong wind shaped by local topography.
The synoptic and fire environment during the October 2017 Wine Country fires are described in Reference [9
]. A trough moved from the Pacific Northwest through the Great Basin on 8–9 October (Figure 2
a,b) creating a regional pressure gradient generating winds first driven south down the Sacramento Valley (Figure 2
c), then southwestward over the coastal and eastern Sacramento Valley mountain ranges (Figure 2
d), encountering a sequence of topographic features with different spatial scales.
Weather station data presented a complex image of the local airflow. Several stations indicated a dramatic increase in wind speeds peaking near 12 a.m.–1 a.m. (all times Pacific Daylight Time (PDT)) on 9 October; for example, Hawkeye (HWCK1), Lake County RAWS1 (LKRC1), and Kenwood (KENWW) produced wind gusts of 36, 22, and 20 m/s, respectively, while others showed a more gradual increase beginning about 8 p.m. on 8 October. While some higher wind speeds and gusts were located on higher elevation stations (Figure 3
, Figure 4
and Figure 5
) (e.g., HWCK1 and LKRC1), not all nearby mountain stations recorded high winds (e.g., ATLC1). In addition, some of the highest wind speeds and gusts were located at RSAC1, on a low hill just above the valley floor, while nearby valley stations at similar positions (EW2362, EW1582, and KENWW) in the lee of a mountain ridge recorded different strengths. Gusts, the peak wind measured in the mean wind sampling period, were approximately twice the mean wind speed. These characteristics suggested speeds varied not only with elevation but with distance along the range, with locations where high winds extended out onto the Santa Rosa valley floor. The strength of both mean and gust winds experienced pulses; determination of the frequency is difficult due to the differing reporting frequency of stations.
Late on 8 October, over a dozen electrical safety incidents were reported in Wine Country, northern coastal ranges, and Butte and Nevada Counties [10
]. While the cause is still under investigation, reports placed the Tubbs fire source ignition at 38.60895° N and 122.62879° W at 9:45 p.m. (http://cdfdata.fire.ca.gov/incidents/incidents_details_info?incident_id=1867
) (Figure 6
). Centroids of satellite-borne Moderate Resolution Imaging Spectroradiometer (MODIS) 1 km resolution active fire detections at 11:24–11:27 p.m. filled the area up to 6.7 km from the ignition location, with additional detections located up to 8.5 km away, perhaps detecting embers that may or may not have ignited additional fires. Nearly coincident data from the 375 m resolution S-NPP/VIIRS instrument at 3:09 a.m. and later MODIS observations at 3:35 a.m. showed the leading edge of fire detections had extended past Highway 101 in Fulton, California, 17.6 km from ignition, suggesting that, if the front did not already slow, the leading edge of the fire spread at an average rate of 3.0 km/h over 5.8 h. The VIIRS 3:09 A.M. observation is the only sub-kilometer resolution mapping data of the fire until its next observation at 2:53 p.m. on 9 October.
Atmospheric airflow regimes are shaped by factors including the wind speed above topographic features, atmospheric static stability, and the feature aspect ratio (height divided by half-width) of terrain, in addition to surface roughness and vertical wind shear. In reaching the area of the Tubbs fire, air traveled over varying three-dimensional terrain and, in response to the evolving pressure distribution, the wind direction and strength changed with time; thus, direct comparison to theoretical studies were not possible. However, the near surface flow regime over topographic features in the area could be estimated from the non-dimensional Froude number, where, for continuously stratified, two-dimensional flow over a feature of height h
, with incoming wind speed U
, and Brunt-Väisälä frequency N
, is given as
In Equation (2),
is the atmospheric potential temperature,
is the acceleration due to gravity, and z
is the height. Using θ
= 303 K, g
= 9.81 m/s2
, and 7 K/km, N
is approximately 0.015 s−1
in a 1.5 km deep layer of statically stable air near the surface (seen in thermodynamic profiles of the atmospheric base state, not shown). A shallow near-surface stable layer of rapid winds was also noted in a previous simulation of the Esperanza Santa Ana-driven wildfire [11
]. A stable layer inhibits vertical growth of the fire plume as well as shaping the airflow regime. In Reference [11
], it produced pulses of high wind speed at topographic features that broke off to travel downstream at 5–7 min intervals.
These perceptible terrain, wind, and atmospheric stability parameters indicated potential similarities with previously studied airflow regimes such as the strong downslope winds of windstorms in other locations (e.g., foehns [12
], Santa Anas [11
], windstorms of the Cascade mountains [14
], Front Range windstorms [16
]), flow over undulating or multiple hills [18
], neutral flow over small hills [20
], and the acceleration of landfalling hurricane flow over hills [21
]) but is not exactly any of them. Stable airflow past mountains can lead to various phenomena in the wake [22
], and near-surface observations may detect lee waves, rotors, and shedded vortices. Previous work has surmised that at sufficiently high speeds, Fr
>> 1, and the flow may behave as though stratification were neutral [23
]. Because the conditions are too complex for analytical solutions, we rely on numerical simulations.
This work examines the structure of topographically-altered airflow within a case of Diablo winds associated with a widespread fire outbreak in the coastal and inland mountains north of San Francisco Bay in October 2017. Because surface weather station data was sparse and largely missed extreme winds that anecdotal reports suggest occurred, the large-scale synoptic environment was changing throughout the event, and the three-dimensional topography was too complex for analytical solutions, numerical simulations refining from mesoscale to microscale were used to examine the flow regime. When two-way coupled with a fire behavior module—a system validated on previous fires in diverse conditions—CAWFE simulations reproduced many aspects of the unfolding fire event. In showing that observed regional Diablo winds accelerated further over broad ridges in the North Bay, varying throughout the day in strength and direction and creating leeward streaks and pulses of higher wind speeds, the simulations provided a coherent framework for understanding the patterns and variability in the surface station data.
Previous work (e.g., Reference [8
]) has described the regional environmental conditions that indicate and induce strong Diablo winds. Here, research simulations with refined resolution gave further insight into the locations and possible mechanisms for the generation of hypothesized extreme winds within these streaks and pulses—not, as might be anticipated, on the highest ridges, but on hills in the lee of the larger ridges, including near the Tubbs fire ignition site, and in bursts that periodically break off and travel downstream from these secondary hills. The airflow pattern showed aspects of previously studied regimes including transient resemblance to a hydraulic jump, but, as the dimensional analysis of flow parameters support, some of the most extreme winds occur over secondary hill crests. Previous conjectures that stable high-speed flow would behave as though it were neutrally-stratified flow over small hills, producing winds peaks at secondary hill crests despite a near-surface stable layer, were largely borne out but that analysis did not predict the production of the periodic wind bursts. A previous numerical modeling study of a Front Range downslope wind storm [17
] produced similar behavior—bursts of high speed wind breaking off in the mountains’ lee—of similar periodicity, but the authors’ attribution of the phenomena to wave breaking does not apply here; a simpler explanation may underlie both.
As these simulations are performed at resolutions more than an order of magnitude finer than numerical weather prediction forecasts, it may be asked whether such features—well beneath what operational forecasts currently can resolve—can be anticipated with modern computational resources. Assertions have been widely made that coupled models cannot be used in a forecast sense because of the computational demands that would be required, while understanding within the meteorological community states that predictability decreases at finer scales as one approaches the boundary layer regime, where key features cannot be predicted deterministically at all [43
]. We show, as a practical disproof, that the coupled system, configured as a generally applicable forecast rather than optimized for this case, can be run four times faster than real time for a large fire at high resolution on a single processor of a workstation, reproducing the important aspects of the flow regime and providing a good reproduction of the accompanying fire growth. Previous work [44
] showed that, provided that the simulation can reproduce winds at hundreds of meters and include feedbacks from the fire on the local winds (a lesser factor in wind-driven fires [11
]), many aspects of fire behavior are understandable and predictable. Here, we additionally show a practical implication: at least using this model, such computations are fast enough even without multiprocessor supercomputers to be used with sufficient resolution and have good skill as a forecast for large wildfires.
Some points remain unsettled in the reproduction or prediction of winds. The features modeled occurred at scales that may lie between or slightly miss observations, preventing direct verification of extreme small-scale motions or winds at secondary peaks. In addition, their size is beneath the scales that operational forecast models can resolve, limiting the ability to reproduce the extreme winds with current operational models and configurations. The temporal frequency (4–7 min) of bursts in the lee of topographic features are difficult to reconcile with reported winds, which are reported (depending on the equipment, network, and sampling frequency) as several minute averages and gusts. In addition, widely used methods, such as the logarithmic wind profile for vertically adjusting results to measurement heights or other heights of interest, rely on assumptions (e.g., that the stratification is neutral and the flow is stationary) that are not strictly met and may not accurately represent the vertical structure of wind speeds in downslope wind conditions.
Additional issues remain in the prediction of fire growth. Some or all of these could account for the differences between modeled and observed spread rates. First, the use of a flaming front to represent fire spread does not include the role of long distance ember spotting. Fire brands can play many roles in wildland fires, from short-distance spotting ahead of the line where they may be overrun by the flaming front before they can develop their own circulations, or the rarer long-distance spotting kilometers ahead of the line, where they might ignite another fire, or as has been described in recent windstorm-driven fires, the fire itself may be described as traveling as a storm of embers. The treatment of spotting in coupled model simulations of landscape-scale events has been a broad-brush attempt to claim the more rapid rates of spread this provides while overlooking the details and distinctions between these roles. For example, short-distance spotting is assumed to be part of the processes moving the flaming front ahead, processes that are parameterized with relationships like Reference [29
]. Long-distance spotting has been treated as resulting from the resolved-scale winds, perhaps with stochastic components to achieve the apparent randomness, though deterministic prediction of the specific location and timing of the occasional ember is not realistic. Observations have not determined critical aspects of ember storms, which present an intermediate case and it is uncertain if they draw air currents that may impact the upcoming fire line or get overrun, having no impact of subsequent fire spread. Fire growth due spotting depends on the ember density, how favorable conditions are for ignition where they land, and the heat flux they produce. Because VIIRS could detect these fires as soon as they reach one to a few meters in size, or even a storm of a sufficient number of embers whether or not they have ignited other materials, spotting complicates the interpretation of the fire extent from active fire detection data, in addition to muddying the interpretation of anecdotal reports on where the fire is. Therefore, in these conditions, it is not clear if variance in simulated fire extent, which means the extent of the flaming front, from the observed fire extent is an error from the location of the actual flaming front, or if detections indicated the extent of embers, which may or may not have been participating in propagation of the fire. Thus, the lag in modeled fire spread in Figure 10
with respect to reports or detections may correctly represent a difference in extent of the flaming front and embers further ahead.
Second, the empirical rate of spread relationships such as that of Reference [29
] have not been validated at extreme wind speeds. Third, initiation of fire growth forecasts can begin when a fire is detected or reported. In this case, which reported ignition occurred near 9:45 p.m., the first detection by the VIIRS data was produced 5 h later at 3:09 a.m., as the first rapid growth period was ending. Fourth, weather models must also make assumptions about the vertical profile of wind in the atmospheric surface layer; though widely used anyway, the assumptions under which these profiles were developed are not supported in stable, topographically-varying, temporally changing conditions. Finally, the leading edge of the simulated fire traveled fast enough to exit the CAWFE domain showing limits in this approach as currently applied for the detection and forecast of very fast-moving fires.
We found that as high speed regional flow of the type called Diablo winds varied in strength and direction, it created flow patterns that resembled previously studied regimes. However, because of the complex multi-scale topography and evolving synoptic situation, that flow contained some surprising aspects such as streaks and bursts in the lee of some topographic features and stagnation downstream of others. Diablo winds alone were not sufficient to explain extreme winds, nor was their general acceleration over broad hills; instead, the regional pattern and mesoscale (2–20 km) acceleration amplified at small hills, under 2 km wide with heights of a few hundred meters over surrounding terrain, creating exceptional peaks of 30–40 m/s over these secondary, lower hill crests; under a transient feature sometimes resembling a hydraulic jump; and in bursts breaking off and traveling downstream at approximately 5–7-min intervals. The acceleration over secondary hill peaks occurred because in very high Froude number flow, stability effects are diminished and the flow behaves much like neutrally stratified flow over a small hill, though the shedding of bursts of high speed air is not anticipated by that simple analysis. The transience in those peaks and the bursts traveling downstream could be consistent with the variability in data in (Figure 3
, Figure 4
and Figure 5
) from local weather stations in the APRSWNET/CWOP network, for which data is available at higher frequency. The cumulative effect of the complex multi-scale airflow was to reproduce fire arrival times, direction, and shape including growth perpendicular to a flank when the fire reaches a topographic feature and draws itself up.
CAWFE coupled weather and fire growth simulations reproduced the fire shape and rapid growth. A detailed retrospective research simulation lagged observed fire line progression, predicting arrival at Santa Rosa 1 h after satellite active fire detections recorded ignitions there. Another CAWFE simulation configured as a forecast at 370 m horizontal grid spacing produced slightly faster arrival in line with observations, with calculations proceeding at 4 times faster than real time on a single computer processor. The significance is that CAWFE forecasts much finer than operational weather forecasts can anticipate extreme winds as well as fire growth. As the magnitude of extreme winds are often a critical predictive metric, this model may have application to understanding and anticipating high-impact fine-scale wind extrema in applications such as anticipating impacts on utility infrastructure and other phenomena, such as Front Range windstorms and landfalling hurricanes.