Sensitivity Simulations of the 30 March 2020 Xichang Wildfire in Southwest China Based on the WRF-Fire Model

1.   Introduction

Every year, wildfires affect millions of people and cause serious damage worldwide. The wildfire not only impacts the environment, damages the ecological structure, and threatens people’s physical health, but also incurs huge cost of disaster prevention and reduction (Jiménez et al., 2018). The disastrous effects can be relieved if the predictive information of fire behavior can be rapidly obtained for prevention. Therefore, it is urgent to develop tools for predicting the wildfire behavior, and provide key supporting information for wildfire prevention, firefighting decision-making, and fire risk assessment.

To date, there have been many wildfire models, which can be categorized into two groups: empirical models and physical models (Sullivan, 2009a, b). Based on substantial observation data, the empirical models establish a set of empirical equations that consider the combined effects of combustion, chemistry, heat transfer, and fluid dynamics, aiming to rapidly predict the spread rate and direction of fire line. In physical models, numerical calculation is carried out based on understanding of the physics of combustion, to reproduce the main process of real combustion in detail.

Due to the different equation sets in these models, the requirements for computational resources are quite distinct. Currently, operational software primarily relies on empirical models for fire initial spread (Rothermel, 1972; Wang, 1983), such as BehavePlus from the United States Department of Agriculture (Andrews, 2014) and FARSITE (Fire Area Simulator) from the Missoula Fire Sciences Laboratory (Finney, 1998). This type of softwares can be installed on common desktop computers and run with exceptionally fast calculation speed. To emphasize the forecast operability, the software does not consider physical process; nonlinear process; the interactions among fire, atmosphere, and terrain; or three-dimensional phenomena (Van Wagner, 1987a, b). Another kind of softwares is based on computational fluid dynamics, such as FIRETEC (Linn et al., 2002) and Wild Fire Dynamics Simulator (WFDS; Mell et al., 2007), which requires deployment on supercomputers for parallel computing. As their calculation time is far longer than real time, they are more suitable for detailed research on the combustion mechanism.

Traditional wildfire prediction models, whether empirical or physical, rely solely on observation data from nearby automatic weather stations as the input information. However, these observation data are zero-dimensional (single point), failing to capture the inhomogeneous near-surface meteorological conditions over complex terrain. Furthermore, wildfires mostly occur in sparsely-populated areas, resulting in extremely scarce observation data. Additionally, the interaction between fire and atmosphere is not considered in traditional models (Pastor et al., 2003). Yet, weather significantly influences wildfires, particularly the wind that plays a leading role in fire spread. In turn, conflagration can also impact weather (Clark et al., 2004; Coen et al., 2013; Lareau and Clements, 2016, 2017). Buoyancy induced by heat produced during combustion can alter near-surface meteorological fields, increasing fire line spread rate and fire intensity, and transporting moisture and heat further beyond the fire area.

In summary, wildfire models that consider physical processes and fire–atmosphere interaction, with improved resolution, prove beneficial for fire behavior prediction (Pastor et al., 2003). However, it is necessary to reach the balance among physical processes, simulation accuracy, and computational costs from the perspective of real-time forecast. With the rapid development of computers in recent years, there is a tendency of developing a new generation of high-resolution wildfire prediction models that consider fire–atmosphere interaction through a two-way coupling of the micro numerical weather forecast model and the fire model.

As a widely-used mesoscale numerical weather forecast model, the Weather Research and Forecasting (WRF) model has a flexible dynamic framework (Skamarock and Klemp, 2008). By dynamical downscaling, the WRF coupled with a large eddy simulation (LES) model (WRF-LES) is established for high-resolution microscale meteorological simulations and forecasts at the sub-hundred-meter level (Liu et al., 2018, 2020, 2022). WRF also contains a fire model, which is a semi-empirical physical model for fire spread (Mandel et al., 2014) based on the level-set method (Muñoz-Esparza et al., 2018). The WRF-Fire can conduct simulations of not only ideal cases, but also real wildfire cases through coupling with the WRF-LES under time-varying weather conditions.

Based on the WRF-Fire, Jordanov et al. (2011) simulated the fire in Harmanli, Bulgaria. Jiménez et al. (2018) reproduced the fire that occurred in the Great Smoky Mountains National Park in Tennessee of the US on 23 November 2016. Both studies indicated a consistency in burned area between the prediction and observation. The predictions of wind, fire, and smoke propagation obtained by the model can greatly reduce the fire losses. Yang and Dong (2020) simulated the evolution of the wildfire on 21 May 2009 that spread from Mongolia into China, and assessed the effect of fire-prevention isolation barrier along the China–Mongolia border.

However, the aforementioned studies have focused primarily on the potential of the WRF-Fire as a simulation tool, where the meteorological conditions driving the wildfire model and the influence of the fire–atmosphere interaction on the fire evolution were not involved. In this study, the wildfire that occurred in Xichang City, Sichuan Province in China on 30 March 2020 (“3·30” Xichang fire hereafter) is simulated by carrying out several groups of experiments and compared with observation. Within these experiments, the performances of the WRF-Fire under different model horizontal resolutions are assessed, and the impact of the fire–atmosphere interaction on the evolution of fire behavior is analyzed.

The remainder of this paper is organized as follows. Section 2 briefly introduces the “3·30” Xichang fire, the associated circulation patterns, and the observation data. Section 3 presents the model settings and experimental design. The simulation results of meteorological fields and fire behavior are provided in Sections 4 and 5. Finally, conclusions and discussion are given in Section 6.

2.   The fire case

2.1   An overview of the “3·30” Xichang fire

Xichang is situated in the hinterland of the Anning River Plain in the West Sichuan Plateau, with an average altitude over 1500 m, reaching 2317 m in the Lushan Mountain. At 1535 BT (Beijing Time) 30 March 2020, a fire took place in the Lushan Mountain scenic spot. The ignition point (27°48′20″N, 102°12′56.03″E, marked with a yellow star in Fig. 1) is located at the ridge along the border between Jingjiu Village and Anha Town. Dense surface vegetation, comprising over 90% shrubs, grasses, and arbors, which consist of eupatorium adenophorum with an average height of 70 cm, arbors with an average height of 80 cm, and pinus yunnanensis with an average height of 100 cm, was found near the ignition point. The fire rapidly spread towards the urban area in the northeast along with the southwesterly wind (reaching 20.9 m s⁻¹ of Grade 9 at 1556 BT).

Download

Larger image

PowerPoint slide

Fig  1.  Burned area (white line) with land cover extracted from the Sentinel-2 satellite. Note that the yellow star in the lower left represents the ignition point.

According to the Survey Report of “3·30” Forest Fire Event in Xichang City (http://www.scsqw.cn/whzh/slzc1/content_49723), the burned area during the entire conflagration reaches 30.4778 km² (white line in Fig. 1), with an economic loss of 97.3112 million yuan. The conflagration experienced the following four stages:

(1) The first stage (from 1547 BT 30 March to 0050 BT 31 March) dealt with fire reporting and initial disposal. The main tasks during this stage were fire investigation and firefighting planning. At this stage, the fire was in the natural development period, with the highest fire intensity and the fastest fire spread. It is worth noting that the local wind direction shifted from southwesterly to northwesterly from 0004 to 0100 BT 31 March. When the rescue and investigation team passed through the Caijiagou Reservoir (the accident place hereafter), the non-fire area in the south was affected by flying fire (sparks or fireballs moving in the air) carried by the northwesterly wind. Consequently, the team became trapped by the fire, resulting in the loss of 19 lives;

(2) The second stage (from 0050 BT 31 March to the early morning 2 April) was collaborative rescue in fighting the fire. At 0900 BT 31 March, the fire weakened and began to spread southward under the steering of the northerly wind. By 2200 BT 31 March, the emergency management command and rescue team officially assembled and initiated the suppression of the five fire lines from four directions: east, west, north, and south. By early morning 1 April, the conflagration had been fully controlled. However, at noon of 1 April, the extinguished fire line in the southeast of the Lushan Mountain resumed and spread towards the Woyun Mountain. The fire was fully extinguished again by the early morning 2 April. Therefore, it is reasonably inferred that the fire was essentially in a natural development period before 2200 BT 31 March;

(3) The third and fourth stages (till 1600 BT 2 April) were to clear remaining smoke points.

As mentioned above, it is generally considered that the fire is in the natural combustion stage before 2200 BT 31 March. Hence, the period from 0800 BT 30 March to 2000 BT 31 March is selected as the simulation period. Except the re-burned Woyun Mountain due to the resuming fire at the noon 1 April, it is assumed that the fire was effectively controlled from the third stage. Therefore, the final actual burned area is selected as the reference to verify the simulation performance. The simulation period is further divided into four stages for study:

Stage 1: Ignition stage (1547 to 1700 BT 30 March);

Stage 2: Northward-spread stage (1700 BT 30 March to 0130 BT 31 March), with the accident time set at 0100 BT 31 March;

Stage 3: Southward-spread stage (0130 to 1100 BT 31 March);

Stage 4: Ember stage (1100 to 2000 BT 31 March), during which the open fire was basically extinguished.

2.2   Synoptic condition

The synoptic condition during the case period is illustrated in Fig. 2 based on the Rapid-Refresh Multiscale Analysis and Prediction System (RMAPS-ST) analysis data with a resolution of 9 km × 9 km at 6-h intervals, provided by the Institute of Urban Meteorology of China Meteorological Administration. The RMAPS-ST assimilates conventional data such as sounding, surface observation, auto weather station (AWS), and ground-based and aircraft reports.

Download

Larger image

PowerPoint slide

Fig  2.  Patterns of atmospheric circulation at (a) 0800 BT, (b) 2000 BT 30 March, and (c) 0800 BT, (d) 2000 BT 31 March 2020. The thick black contours represent 500-hPa height (gpm), full barb (m s⁻¹) at 700 hPa represents 4 m s⁻¹, the color shaded is the 850-hPa temperature (°C), and the black boxes represent target area of the “3.30” Xichang fire, which is also the model area of d03 in this study.

Due to the high elevation of the West Sichuan Plateau, at around 3000 m, the selection of the 700-hPa wind field in Fig. 2 better represents the local background wind and is also closer to the surface. The choice of using the 850-hPa temperature field is due to the more prominent warm center at 850 hPa.

At 0800 BT 30 March, the fire area was situated in front of the eastward-moving shallow trough at 500 hPa and an obvious wind shear at 700 hPa. At night (2000 BT), there was an obvious warm center at 850 hPa over the southern part of the fire area, reaching as high as 25°C. Affected by the southwesterly wind at 700 hPa, there was a noticeable warm advection over the fire area. As mentioned above, the fire was ignited at 1535 BT and quickly spread northeastward under the influence of the 700-hPa southwesterly wind. Meanwhile, the upper-level trough further deepened, and the 700-hPa wind shear began to pass through the fire area. This indicates a wind shift from southerly to northerly over the fire area at night as the weather systems moved eastward. The synoptic condition corresponds well with the sudden change in wind direction from southerly to northerly in the accident place around 0000 BT 31 March.

At 0800 BT 31 March, an upper-level trough at 500 hPa and a weak short-wave trough at 700 hPa had passed over the fire area, where the northwesterly wind was strong, causing the fire to spread southward. At 2000 BT 31 March, the fire area was located behind the trough, further promoting the northeastward spread of the fire line under the intense southwesterly warm advection.

2.3   Observation data

To quantitatively evaluate the simulation effect, the real burned area is extracted based on the Sentinel-2 Level-1C (L1C) data (Sentinel-2A on 25 March before the conflagration and Sentinel-2B on 9 April after the conflagration). Based on L1C, after performing radiometric calibration, atmospheric correction, rational polynomial coefficients (RPC) orthorectification and resampling, the 10-m high-resolution multi-band data are obtained. The burned area is then extracted by the burned area index (BAI) method (Sun et al., 2019). The BAI is calculated based on the spectral distance between each pixel and a reference spectral point. The red and near-infrared reference reflectance values are 0.06 and 0.1, respectively, which tend to emphasize the charcoal signal of burned areas. The calculation formula is as follows:

where $\rho_{\mathrm{R}}$ and $\rho_{\mathrm{NIR}}$ represent the red and near-infrared reflectance, respectively. By comparing the average BAI values of surface vegetation before and after the conflagration, the detailed burned area is ultimately extracted (black line in Fig. 1). The extracted burned area measures 30.2529 km², slightly smaller than the official area of 30.4778 km².

In addition, to verify the simulation effect of surface meteorological fields, the simulation results are compared with data from the Xichang AWS (27°54′0″N, 102°16′12″E; station number 56571, marked as a blue dot in Figs. 3b, c), situated in the northwest of the fire area.

Download

Larger image

PowerPoint slide

Fig  3.  Computational domain for WRF-Fire with the terrain elevation and state boundaries. The red pentagram represents the fire point, the blue dot represents the Xichang AWS, and the black triangles represent densely populated area, such as school and historical scenic spot.

3.   Model setting and experiment design

3.1   WRF-Fire setup

To simultaneously represent scales from the synoptic to turbulent eddies, the LES version of the WRF (WRF-LES, version 4.3.1) model system is configured with a five-domain one-way online nested setup (Table 1). There are three mesoscale domains: d01 (Δx = 13.5 km), d02 (Δx = 4.5 km), and d03 (Δx = 1.5 km), as well as two microscale domains: d04 (Δx = 0.5 km) and d05 (Δx = 0.1 km), with a refinement ratio of 5. Notably, d05 covers the burned area of the “3·30” Xichang fire.

Table  1.  Domain size, grid spacing, and timestep in the WRF-Fire model

Download as CSV

Domain	Atmospheric model nx × ny × nz	Atmospheric model Δx (km)	Model top (hPa)	Time step (s)	PBL scheme	Coupled with fire	Fire model nx × ny × nz	Horizontal grid spacing ratio	Fire model Δx (m)
d01	121 × 121 × 50	13.5	50	45	YSU	No	–	–	–
d02	121 × 121 × 50	4.5	50	15	YSU	No	–	–	–
d03	121 × 121 × 50	1.5	50	5	YSU	No	–	–	–
d04	121 × 121 × 50	0.5	50	5/3	LES (TKE-1.5)	Yes	2420 × 2420 × 50	1 : 20	25
d05	221 × 221 × 50	0.1	50	1/3	LES (TKE-1.5)	Yes	884 × 884 × 50	1 : 4	25

 | Show Table

DownLoad: CSV

The main distinction of WRF-LES from the commonly used WRF in mesoscale is its high horizontal resolution has reached the microscale. As a result, traditional one-dimensional (1D) planetary boundary layer (PBL) schemes, based on Reynolds-Averaged Navier-Stokes (RANS) equation, are not applicable. In addition, 1D TKE (turbulent kinetic energy) closure scheme only considers vertical exchange, while horizontal exchanges and three-dimensional (3D) effects such as horizontal shear production are entirely neglected over mountainous terrain (Goger et al., 2018, 2019). Hence, in this study, within the inner domains d04 and d05 of WRF-LES, the PBL scheme is turned off and the boundary layer turbulence is simulated by LES. LES can explicitly resolve the most energetic of 3D turbulent eddies, and the turbulence closure in the LES may be considered as an alternative to PBL schemes (Cuxart, 2015). In this study, the LES used 1.5-order TKE closure model (TKE-1.5) to parameterize sub-grid scale (SGS) flux. In addition to utilizing a length scale and a model coefficient, it also computes the SGS TKE as a prognostic variable to obtain the eddy viscosity (Lilly, 1966).

Furthermore, based on the fire module in the WRF, the two LES-scale simulation areas of d04 and d05 are both coupled with the fire model to build a WRF-Fire system. The horizontal resolution of the fire model is 25 m × 25 m, and the horizontal grid spacing ratios are 1 : 20 and 1 : 4 for d04 and d05, respectively. The specific settings are shown in Table 1, and the simulation ranges are shown in Fig. 3. In the model, 17 vertical layers are set below 1 km. Take d05 as an example: the height of the first model vertical level is approximately 25 m AGL near the surface (${\Delta }{z}$/${\Delta }{x}$ = 0.25); below 200 m,${\Delta }{z}$ is uniformly set as 31 m (${\Delta }{z}$/${\Delta }{x}$ = 0.31); from 200 m to 1 km, ${\Delta }{z}$ ranges from 42 m to approximately 97 m (${\Delta }{z}$/${\Delta }{x}$ = 0.42–0.97).

The physical schemes utilized in the model from d01 to d05 include the Thompson scheme (Thompson et al., 2008) for microphysical processes, the rapid radiative transfer model (Iacono et al., 2008) for longwave and shortwave radiation options, the Noah land surface scheme (Tewari et al., 2004), and the revised mesoscale model version 5 (MM5; Zhang and Anthes, 1982) for near-surface layer processes. In addition, the Kain–Fritsch cumulus parameterization scheme (Kain, 2004) is only used in d01 and d02. The Yonsei University (YSU) PBL scheme is applied for boundary layer processes solely in d01–d03, with the modified subgrid-scale terrain parameterization scheme opened in the YSU scheme (Liu et al., 2019). In d04 and d05, the YSU PBL scheme is turned off and the LES (TKE-1.5) is turned on. The single point (27°48′20″N, 102°12′56.03″E, marked by red dots in Figs. 2c, 3b) ignition is used in d04 and d05, with the ignition radius set at 30 m (> ${\Delta }{x}$ = 25 m), a duration of 10 s, and an ignition speed of 0.1 m s⁻¹ (default).

The initial fields and lateral boundary conditions are derived from the Global Ensemble Forecasting System (GEFS) from the NOAA, which is at 6-h intervals with horizontal grid spacing of 0.5°. The elevation data are from the Shuttle Radar Topography Mission 1 (SRTM1; http://www2.jpl.nasa.gov/srtm/) datasets with a resolution of 30 m. The landuse data are obtained from the Global Land Cover product with Fine Classification System in 30 m for 2020 (GLC_FCS30-2020; https://data.casearth.cn/sdo/detail/5d904b7a0887164a5c7fbfa0). This dataset contains 29 land cover types, which are mapped with the 27 land cover types from the United States Geological Survey (USGS) to facilitate data reading and writing for the WRF.

Due to the absence of basic datasets of surface fuel types in China, a mapping method is applied to obtain the dataset of surface fuel types in this study. Specifically, the 29 land cover types from the GLC_FCS30-2020 are mapped to the 13 fuel types (Anderson, 1982) to obtain the 30-m dataset. For instance, the land cover in d05 mainly consists of seven types: low grass, shrubs, dwarf forests, dense forests, bare areas, cities, and water bodies, which respectively correspond to type 1 (low grass), type 5 (shrubs), type 7 (dwarf forests), type 8 (dense combustible waste such as fallen leaves and wood), type 14 (non-combustible substances), type 14, and type 14 as classified by Anderson (1982), as shown in Fig. 4. In the actual simulation, the fuel depth of type 1 (low grass) is adjusted from the original default value of 0.305–0.150 m, while other fuel parameters remain unchanged.

Download

Larger image

PowerPoint slide

Fig  4.  High-resolution landuse maps and fuel categories in d05. The red star represents the ignition point, and the blue dot represents the Xichang AWS.

3.2   Sensitivity experiment design

As shown in Table 2, two groups of sensitivity experiments are designed to compare the impacts of meteorological conditions with different resolutions (100 and 500 m) and the fire–atmosphere interaction [feedback(fb) or no feedback(nfb)] on the evolution of fire behavior. In comparison to 100nfb, 100fb takes into account the bidirectional feedback between the atmosphere and fire, meaning that the sensible heat, latent heat, and aerosols generated during the combustion are synchronously transmitted to the meteorological model for further integration. The other two control experiments of 500nfb/500fb denote the experiments performed under the horizontal resolution of 500 m in d04 without/with the fire landuse–atmosphere interaction.

Table  2.  Experiments setup

Download as CSV

Experiment type	Experiment name	Atmospheric model Δx (m)	Nesting ratio	Fire model Δx (m)	Fire–atmosphere interaction (feedback)
Control	100nfb	100	1 : 4	25	No
Compare	100fb	100	1 : 4	25	Yes
Compare	500nfb	500	1 : 20	25	No
Compare	500fb	500	1 : 20	25	Yes

 | Show Table

DownLoad: CSV

For simplicity, the experiments of 100nfb and 100fb are referred to as 100-m group, while the experiments of 500nfb and 500fb are henceforth termed 500-m group. Additionally, the experiments of 100nfb and 500nfb are denoted as the nfb group, and the experiments of 100fb and 500fb are designated as the fb group. The initial time of all four tests is 0200 BT 30 March, with the first 6 hours as the spin-up time. The actual simulation period spans from 0800 BT 30 March to 1200 BT 31 March, totaling 36 h, with an output interval of 5 min.

4.   Analysis and verification of meteorological condition

4.1   Time series of meteorological elements at Xichang AWS

Figure 5 exhibits the observed and simulated 2-m temperature (T₂), 2-m specific humidity (q₂), and 10-m wind speed (V₁₀) and direction (WD₁₀) from four experiments at Xichang AWS. Table 3 shows the mean error (ME) and root-mean-square error (RMSE), with the calculation formulas outlined as follows:

Download

Larger image

PowerPoint slide

Fig  5.  Comparison of the time series of observed and simulated (a) T₂, (b) q₂, (c) V₁₀, and (d) WD₁₀ from four experiments at Xichang AWS. The color shading emphasizes significant change of meteorological elements at Xichang AWS 2 h prior to the accident time.

Table  3.  Mean error (ME) and root-mean-square error (RMSE) of T₂, q₂, V₁₀, and WD₁₀ of four experiments at Xichang AWS

Download as CSV

	100nfb			100fb			500nfb			500fb
	ME	RMSE		ME	RMSE		ME	RMSE		ME	RMSE
T2	−0.61	1.73		−0.29	1.60		−3.37	2.50		−3.38	2.52
q2	−5.71	6.78		−5.51	5.98		−5.79	11.96		−5.79	12.05
V10	0.50	1.37		0.34	1.50		2.49	1.37		2.45	1.37
WD10	−13.79	56.69		−7.60	53.08		32.50	61.38		32.30	61.75

 | Show Table

DownLoad: CSV

where $P$ and $O$ represent the predicted and observed values, respectively, and $n$ is the number of samples.

Figure 5 and Table 3 indicate that the near-surface wind, temperature, and specific humidity simulated by the control test 100nfb are generally consistent with the observations at Xichang station. The RMSE is within 2°C for T₂, 6.78 g kg⁻¹ for q₂, 1.37 m s⁻¹ for V₁₀, and 56.69° for WD₁₀. The most notable difference between 100nfb results and observations occurs from 2100 BT 30 March to 0100 BT 31 March, during the northward propagation and strengthening of the fire. There was a sudden and significant increase in T₂ at the Xichang station, accompanied by a sharp decrease in q₂ and V₁₀. Meanwhile, WD₁₀ shifted from southerly to northwesterly. This phenomenon is not well captured by the 100nfb simulation. In contrast, for 100fb, the results are almost identical before the ignition, but gradually begin to show differences after ignition. 100fb considers the atmosphere–fire interaction, and its results are much better than 100nfb, although with a smaller variation amplitude and a delayed transition time. These results indicate that the sudden change of meteorological elements is caused by the fire–atmosphere interaction, but the effect of fire on the atmosphere is still underestimated in the model.

For 500-m group, the RMSEs of the simulated T₂ and q₂ are both larger than those of the control experiment. The curves of the V₁₀ time series are smooth and even, failing to capture the characteristics of microscale wind speed fluctuations. WD₁₀ remained southerly during the entire simulation period. Besides, the 500fb that considers the atmosphere–fire interaction cannot reproduce the wind, temperature, and humidity shifts observed in the 100fb, and there are minimal differences between 500fb and 500nfb after the ignition. This can be attributed to several factors. First, Xichang station is distantly located from the fire area, and thus experiences less impact. The same goes for 100-m group. For instance, after 0100 BT 31 March, the spread of the fire line turned from northward to southward under the wind shift. At this point, Xichang AWS is located upwind of the fire site, leading to a significant reduction in the influence of the fire, consequently diminishing the disparities between 100fb and 100nfb. Second, the fire area cannot be well described with a coarse model resolution of 500 m, where the heat feedback from the combustion within the grid to the atmospheric model is weakened. These results show that it is necessary to improve the resolution of atmospheric model as much as possible to fully consider the fire–atmosphere interaction, and to obtain more accurate fire behavior.

4.2   Meteorological fields

To further analyze the impact of horizontal grid spacing and the fire–atmosphere interaction on the simulation of meteorological fields, Fig. 6 shows the boxplot of the T₂ difference among different experiments for all grids points during the entire simulation period. The temperature difference between fb and nfb groups serves as an indicator of the fire–atmosphere interaction, while the disparity between the 100- and 500-m groups (with simulation results interpolated onto the 100-m grid) reflects the impact of horizontal resolution.

Download

Larger image

PowerPoint slide

Fig  6.  Boxplot of T₂ difference among each experiment.

Figure 6 reveals that the experiments with the atmosphere–fire interaction have a warming effect on T₂ in the simulation area. Specially, compared to 100nfb, T₂ of 100fb increases by 1.2°C on average, with the maximum warming amplitude reaching 7.9°C. However, the T₂ difference between 500nfb and 500fb only increases by 0.6°C on average, with the maximum warming amplitude at 2.4°C. As a result, the warming effect simulated by 100-m group is much higher than that of 500-m group. The variable T₂ simulated by the 100nfb are lower than those by the 500nfb. This is due to the model terrain smoothed in the 500nfb, resulting in lower altitude and higher temperature on a grid. It is also indicated that a finer horizontal grid spacing in the atmospheric model can better describe the atmosphere–fire interaction.

Figure 7 shows the instantaneous horizontal distributions of the T₂ differences between the fb and nfb groups shortly after the ignition (1745 BT 30 March), at the accident time (0100 BT 31 March), and those of the maximum and minimum T₂ differences during the entire simulation. The heat released directly during combustion is carried to the surrounding of the burned area by local circulation, ultimately influencing the surface temperature. Initially (Figs. 7a, b), the fire–atmosphere interaction has the greatest impact near the fire front, with the maximum increase found in T₂. Besides, T₂ also increases in the long and narrow airflow in the downwind area of the ignition point, while other areas have not yet been affected by the fire. In comparison with the 500-m group, the areas where T₂ is affected by the fire are much longer, narrower, and more concentrated in the 100-m group.

Download

Larger image

PowerPoint slide

Fig  7.  Horizontal distributions of T₂ difference between fb and nfb groups at (a, b) 1745 BT 30 March and (c, d) 0100 BT 31 March, as well as the (e, f) maximum and (g, h) minimum T₂ differences during the simulation.

At the accident time (Figs. 7c, d), it is evident that the large-value area of T₂ difference between 100fb and 100nfb corresponds closely to the actual location of fire line and also the spread direction along the downwind. The instantaneous 10-m wind fields simulated by the four experiments at the accident time (Figs. 8c, d) reveal that the wind direction of 100-m group has shifted from southwesterly to northwesterly, then to northerly at the accident place, while 500-m group failed to reflect this shift, remaining southwesterly. Both the 100nfb and 100fb clearly indicate the risk of wind direction change at accident place. These results reveal the importance of high-resolution forecasting over complex terrain areas, as it can provide accurate information of local meteorological field for disaster prevention and mitigation.

Download

Larger image

PowerPoint slide

Fig  8.  Instantaneous horizontal distributions of simulated 10-m wind field at (a, b) 1605 BT 30 March, (c, d) 0100 BT 31 March, and (e, f) 2000 BT 31 March. The black solid line represents terrain height (m), and vectors represent wind field (m s⁻¹).

According to the maximum (Figs. 7e, f) and minimum T₂ differences (Figs. 7g, h), it is clear that topographic resolution can greatly influence the degree of interaction between fire and the atmosphere. With low topographic resolution, the interaction between fire and the atmosphere has minimal impact on the simulation results. It is only when the topographic resolution is increased that the significant impact of the interaction between fire and the atmosphere becomes apparent. This also indirectly suggests that lower topographic resolution will diminish the model’s ability to simulate vertical motion.

In Fig. 8, the contour lines of terrain show that the terrain is flat in the 500-m group, resulting in a more uniform distribution of the wind field across the entire area. In contrast, the 100-m group can better describe the terrain characteristics of the Lake Qionghai–Mountain Lushan region, leading to the display of detailed microscale local circulation structures in the instantaneous wind fields.

The consideration of the fire–atmosphere interaction results in different impacts on the simulations. In the 500-m group, no significant difference is found on the simulated wind field between fb and nfb experiments, with the wind vectors almost overlapping (Figs. 8b, d, e). While for the 100-m group:

(1) In the ignition stage, slight differences are evident at the ignition point and its downwind region between fb and nfb experiments (Fig. 8a), and the range of wind differences corresponds to the large-value area of T₂ difference (Fig. 7a);

(2) At the accident time (Fig. 8c), a much larger difference in the wind direction is observed between the fb and nfb experiments of the 100-m group, especially in the downwind region of fire area, where the wind field even presents a contrary direction. This phenomenon can be attributed to the updrafts induced by the fire area, leading to the converges of the wind field towards the fire center, accompanied by a noticeable increase in wind speed;

(3) In the ember stage (Fig. 8c), significant differences are found across the entire wind fields between the 100fb and 100nfb experiments with the fire line spreading towards different directions. The combined effects of combustion and terrain leading to the generation of abundant microscale turbulent eddies after considering the fire–atmosphere interaction.

It is worth noting that both the 100-m group and the 500-m group have captured the northwesterly flow and wind direction change. The difference between the simulation results of the two groups is mainly caused by the topography. As the 500-m group uses a coarser topography, resulting in smoother terrain and weaker mountain blocking effects, the simulated southern wind is stronger, while the northern wind is weaker. This leads to a northward shift in the location of the wind convergence line, causing the simulated results to move southward more slowly compared to the 100-m group.

Figure 9 shows the boxplot of simulated wind speed and direction differences for the four experiments at each grid. In the 100-m group, after considering the atmosphere–fire interaction, the simulated average V₁₀ increases by 3.9 m s⁻¹, with the maximum increase reaching 12.2 m s⁻¹. However, in the 500-m group, the simulated average V₁₀ increases by 0.8 m s⁻¹, with the maximum increase being only 2.6 m s⁻¹. In terms of wind direction, the absolute values of the WD₁₀ difference simulated by the 100-m group at 25% and 75% percentiles both reaches 120°. For the 500-m group, the fire–atmosphere interaction has little influence on wind direction, as the average difference is closed to zero and the absolute values at 25% and 75% percentiles are relatively lower. These results once again indicate that the atmospheric model with finer horizontal resolution can better reflect the fire–atmosphere interaction.

Download

Larger image

PowerPoint slide

Fig  9.  Boxplots of (a) V₁₀ and (b) WD₁₀ difference among each experiment.

5.   Analysis and verification of fire behavior

5.1   Fire area

Figure 10 displays the simulated and observed burned areas at 1745 BT 30 March (ignition stage), 0100 BT 31 March (accident time), 1100 BT 31 March (southward spread stage), and 2000 BT 31 March (ember stage). It can be seen that the simulation by 100nfb on burned area is relatively more accurate, covering the fire field except the southeast area (i.e., the Woyun Mountain where the combustion began after noon on 1 April). The simulated areas and shapes of burned area by the four experiments are generally similar. However, the fire area simulated by the 100-m group are slightly larger than those by the 500-m group, primarily concentrated in the southeast of the fire field. During the ignition stage, the fire lines in both the 100- and 500-m groups spread northeastward, with higher spread speed found in the 100-m group. At 0100 BT 31 March, the wind shifts to a northerly direction and the fire line spreads southward in the 100-m group, while the fire line still spreads northward along the background southwesterly wind in the 500-m group. In the southward spread and ember stages, the difference between the two groups further enlarges, with a larger fire area in the southeast for the 100-m group. After considering the fire–atmosphere interaction, the differences in simulated burned area between the fb and nfb experiments are relatively smaller.

Download

Larger image

PowerPoint slide

Fig  10.  Simulated fire area at (a) ignition stage (1745 BT 30 March), (b) northward-spread stage (0120 BT 31 March), (c) southward-spread stage (1100 BT 31 March), and (d) ember stage (2000 BT 31 March) with burned area.

To quantitatively evaluate the simulation effect on the fire area, the simulated burned area is divided into three parts: the overlapping area (a), which represents the part where the simulated and observed areas coincide; the over-simulated area (b), which is burned only in the simulation; and the under-simulated area (c), which is burned only in the observation. Based on the pixel numbers and areas a, b, and c, the KAPPA coefficient (KC; Congalton, 1991) and spatial correlation coefficient (SC; Legendre and Legendre, 1998) are commonly used in fire spread model assessment. The KC is the result of performing a KAPPA analysis (Cohen, 1960), which is mainly used to describe the conformity between the simulation and the observation, with the calculation formula as follows:

where $N$ is the total pixel number of confusion matrix, $r$ is the number of rows for confusion matrix, $x_{ij}$ is the pixel number in row $i$ and column $j$ of the confusion matrix, $x_{i+}$ is the pixel number in areas a, b, and c of the simulated confusion matrix, and ${X}_{+j}$ is the pixel number in areas a, b, and c of the observation. The SC coefficient is used to characterize the correlation between the simulation and the observation, with the calculation formula as follows:

Table 4 shows that the KC values of simulated fire area from the four experiments range from 0.56 to 0.59, while the SC values range from 0.52 to 0.59. These values indicate that the simulation effects of the four experiments on the fire area are similar. Specifically, the simulated overlapping area in the 100fb is the largest, suggesting a slightly superior performance compared to the others. According to a previous assessment on the forest fire behavior simulation model in the forest area of Southwest China (Zhao et al., 2017), KC values generally fell between 0.30 and 0.43, with SC values between 0.39 and 0.50 when using Prometheus (Tymstra et al., 2010) and FARSITE (Finney, 1998) fire model. These research results have proven that the WRF-Fire can well simulate the fire area, which has a great value in future fire prevention and control.

Table  4.  Evaluation of the simulation accuracy of fire area from each test

Download as CSV

Name	Simulated total area (km2)	Overlapping area (km2)	Over-simulated area (km2)	Under-simulated area (km2)	Observed total area (km2)	KC	SC
100nfb	39.48	16.09	8.65	14.74	3025	0.58	0.58
100fb	40.67	16.96	8.79	14.92		0.59	0.59
500nfb	37.98	13.43	4.97	19.58		0.56	0.52
500fb	37.88	13.65	5.24	18.99		0.57	0.53

 | Show Table

DownLoad: CSV

5.2   Fire spread rate

Besides the fire area, the rate of fire spread is also one of the important elements reflecting fire behavior. Figure 11 shows the horizontal distributions of the maximum spread rate at fire front and the difference between fb and nfb experiments at each grid. Figures 11e and 11f reveal that the simulated meteorological fields are relatively flat in the 500-m group due to the coarse grid spacing of the atmospheric model, resulting in the homogeneous distribution. However, the spread rate simulated by the 100-m group (Figs. 11a, c) varies greatly at different locations, with the maximum speed occurring on the windward slope and the minimum speed on the leeward slope. The differences in distribution (Figs. 11c, d) illustrate that the fire–atmosphere interaction has a more pronounced impact on the fire spread rate in the 100-m group, resulting in increased front spread rate on the windward slope and decreased on the leeward slope.

Download

Larger image

PowerPoint slide

Fig  11.  Horizontal distributions of maximum spread rate at fire front of each test and their difference. (a) 100nfb, (b) 500nfb, (c) 100fb, (d) 500fb, (e) 100fb−100nfb, and (f) 500fb−500nfb.

Figure 12 further shows the probability density distribution of the spread rate at fire font in each grid of each experiment, with the average and maximum rate shown in Table 5. During the entire combustion period, the highest proportion of simulated spread rate is less than 0.1 m s⁻¹ among the four experiments. Specifically, the proportions of spread rate less than 0.05 m s⁻¹ simulated by the nfb experiments are higher than those by the fb experiments. Besides, the proportion rapidly reduces as the spread rate increases. The proportion of spread rate less than 0.1 m s⁻¹ (0.07 m s⁻¹ simulated by nfb experiments) in the 100-m group is higher than that in the 500-m group. Therefore, the average fire head speed simulated by the 100-m group is relatively lower, but the maximum spread rate in the 100-m group is larger than that in the 500-m group. The simulation results of the 100-m group (Table 5) show that the average spread rate in the 100nfb is 0.08 m s⁻¹, and the maximum spread rate in the 100nfb is 0.84 m s⁻¹. The average spread rate in the 100fb is 0.08 m s⁻¹, and the maximum spread rate in the 100fb is 0.97 m s⁻¹. Compared with the 100nfb, the average and maximum spread rate in 100fb respectively increase by 6.61% and 15.48% after considering the fire-atmosphere interaction. Compared with the 100-m group, the average spread rate in the 500-m group increases by 32.25% (25.58%) in nfb (fb) experiment. Besides, the maximum rate in the 500-m group is slightly smaller than that in the 100-m experiments, amounting to 98.81% (91.75%) of the maximum rate in the 100nfb (100fb) experiments. In summary, the high-resolution atmospheric model can fully consider the atmosphere–fire interaction, which significantly impacts the spread rate at fire front.

Download

Larger image

PowerPoint slide

Fig  12.  Probability density distribution of spread rate at fire front in every grid of each experiment.

Table  5.  Average and maximum of spread rate at fire front of each experiment

Download as CSV

Experiment	100nfb	100fb	500nfb	500fb
Average rate (m s−1)	0.08	0.08	0.11	0.12
Maximum rate (m s−1)	0.84	0.97	0.83	0.89

 | Show Table

DownLoad: CSV

6.   Conclusions and discussion

In this study, based on a high-resolution atmosphere–fire coupling model—the WRF-Fire, the meteorological fields and fire behavior in 36-h period of the “3·30” Xichang fire are simulated. Four numerical experiments are designed to analyze the impacts of varying horizontal grid spacing and the fire–atmosphere interaction on meteorological fields and fire behavior simulation compared with the observation. The main conclusions are as follows.

(1) The fire field is located in the complex mountain area, and the 100-m experiments can better describe the local wind field and associated microscale wind speed fluctuation compared to the 500-m group. The 100-m group also provides a more accurate simulation of near-surface temperature, humidity, and wind field. This can be attributed to several factors. First, the high-resolution experiment allows for better representation of the terrain features. Second, the high-resolution grid is able to magnify physical processes in the boundary layer. Third, the LES is capable of resolving the most part of atmospheric turbulence motion on the 100-m scale, but failed in 500 m.

(2) A high-resolution atmospheric model can better reflect the fire–atmosphere interaction. For surface meteorological fields, the fire–atmosphere interaction directly causes an increase in temperature, elevated wind speeds, and converging to the fire lines, and indirectly affects local circulation in the downwind region far from the fire field. In this study, for 100-m (500-m) group, the fire–atmosphere interaction causes an average increase of 1.2°C (0.6°C) in T₂, with the maximum increase reaching 7.9°C (2.4°C). Besides, the average V₁₀ increases by 3.9 m s⁻¹ (0.8 m s⁻¹), with the maximum increase reaching 12.2 m s⁻¹ (2.6 m s⁻¹).

(3) The KC values for the simulated fire area in all four experiments range from 0.56 and 0.59, while the SC values fall between 0.52 and 0.59, indicating satisfactory performance. In the 100-m group, the distributions of spread rate at fire front are more inhomogeneous than those in the 500-m group, with smaller average spread rate and larger maximum rate. The maximum spread rate is observed on the windward slope, while the minimum is on the leeward slope. These differences further increase when considering the fire–atmosphere interaction. Overall, for real case over complex terrain area, the influence of heterogeneous land surface factors (i.e., terrain, land-use, and model grid spacing) on fire behavior is much greater than the atmosphere–fire interaction.

In summary, the results demonstrate that the WRF-Fire presents significant advantages as a wildfire simulation tool, providing a new and feasible approach for fire prediction. However, the consideration of the fire–atmosphere interaction in this model is still insufficient. In addition, the dataset of surface fuel types distribution and the table of fuel property parameters is still not accurate enough. Therefore, it is imperative to promptly establish a dataset of fuel types with detailed spatiotemporal distributions nationwide (at least in fire-prone districts), where the parameters of the basic attributes of typical fuel types (fuel depth, the initial mass loading of surface fuel, fuel surface-area-to-volume-ratio, fuel moisture, etc.) are calibrated. Moreover, in future, it is urgent for meteorological, forestry, and fire department to conduct a jointly fire ignition experiments and share observation data, which are crucial for building and optimizing wildfire models.