Ngoring Lake (surface area of 610 km²) and Gyaring Lake (surface area of 526 km²) are located in the source region of the Yellow River (Li et al., 2017). The average altitude of the lake surfaces is 4,274 m above sea level (as shown in Figure 1). Ngoring Lake and Gyaring Lake are large freshwater lakes in the Tibetan Plateau. Around the lakes, the vegetation type is alpine meadow with a height of 5–10 cm. The vegetation coverage ranges from 0.55 to 0.75 (Li and Li, 2015), and the leaf area index is approximately 1.2–1.4 m² m⁻² in summer (Li et al., 2021). The albedo of the lakes and alpine meadow averaged over 12:00–14:00 local solar time (LST) in summer is 0.04 and 0.2, respectively. The grassland eddy flux tower site (97°33′16″E, 34°54′51″N), standing on a flat underlying surface, is located 1.5 km west to Ngoring Lake.

Numerical models play an important role in weather simulation and forecasting (Wang and Yu, 2013; Wu et al., 2014; Zhou et al., 2014). The WRF model was used to simulate the meteorological conditions from 00:00 LST on May 4, 2013 to 24:00 LST on September 30, 2013 in the basin of Gyaring and Ngoring lakes. Table 1 summarizes the experimental parameters of the WRF model. A double one-way grid nesting technique was used. The results were output every hour for both the first and second domains of the WRF model. The model contained 30 layers of atmosphere in the vertical direction. The pressure at the top layer of atmosphere was set to 50 hPa. In addition to a control experiment, a sensitive experiment on the same area with the Gyaring and Ngoring lakes replaced by grassland surface was conducted to analyze the effects of the lakes on the air temperature, local wind field, near-surface energy budget, and precipitation in the lake area.

In this study, we used a backward water vapor tracing method based on the Lagrangian trajectory method (Brubaker et al., 2001; Dirmeyer and Brubaker, 2007; Harding and Snyder, 2012). The hourly output data from the WRF model were analyzed. The specific analysis method is described as follows: first, the period from May 4, 2013 to September 30, 2013 was divided into 30 5-day window intervals. For each window, 1,000 air parcels were randomly released at the grid point where precipitation occurred (total precipitation for each of the 5 days) to comprehensively consider the uncertainties of backward tracing at various longitudes, latitudes, and heights. The probabilities of the spatial heights of the released 1,000 air parcels were determined for a specified longitude and latitude within this grid point by the vertical specific humidity profile (e.g., if 70% of the humidity is at the bottom layer of the model, then there is a 70% probability that an air parcel is released to the bottom layer), ensuring relatively more air parcels in high humidity layers. Then, for each released air parcel, its horizontal and vertical positions at 1 h before the current moment were calculated. The horizontal position was determined by the wind vector at the position of the air parcel at the current moment. The vertical position was determined by the previous moment when the potential temperature in the layer was the same as that at the initial position where the air parcel was released, i.e., each air parcel moved vertically along the constant potential temperature line. The contribution of evapotranspiration from each position where an air parcel passed to the corresponding grid point was calculated. The percentage of contribution was calculated by dividing the evapotranspiration at a position where an air parcel passed by the total precipitation in the vertical layer where the position was located. Each time there was a contribution from a position passed by an air parcel to the total precipitation at the current grid point in the current window, this contribution was deducted from the total precipitation. When the sum of local contributions equals the total precipitation at the current grid point in the current window, backward tracing is completed, and all the local sources for precipitation are considered to have been traced. Water vapor tracing was terminated when an air parcel reached the boundary of the modeled area or the tracing duration exceeded 7 days. Finally, the precipitation contributed by the evapotranspiration from each surface type was calculated, and this precipitation was referred to as surface-induced precipitation.

This section focuses mainly on the short-term climatic characteristics of the study area. Thus, the average simulated and observed values obtained at the grass site were compared. As shown in Figure 2, the model well captured the overall characteristics of the diurnal cycles of sensible and latent heat fluxes and 2 m air temperature with the R ² ranging from 0.89 to 0.99. However, the model overestimated the daytime sensible heat flux by 40 W m⁻². In addition, the 2 m air temperatures between 00:00 LST and 09:00 LST during the simulation period were underestimated by the model.

The WRF model did not well capture the diurnal cycles of wind speed and precipitation and overestimated the wind speed and precipitation (Table 2) during June, July, and August with a mean overestimation of 0.58 m s⁻¹ and 0.05 mm h⁻¹, and such overestimation mainly occurred in nighttime (Figure 2). The observed wind direction showed clear lake–land breeze wind shift, while the WRF simulated wind shift was not as strong as the observation.

There is a substantial change of the latent heat flux in the original lake area after removal of the lakes (Figure 3A). The results from the experiment with the lakes show the latent heat peaks in September (196.0 W m⁻²). After removal of the lakes, the latent heat in September decreases to 41.0 W m⁻², which is the lowest among all months; in addition, the latent heat decreases substantially in other months. The average latent heat of the lakes between May and September simulated by the experiment with the lakes is 155.7 W m⁻². Without the lakes, the average latent heat decreases by 58.5% in 64.7 W m⁻². From a perspective of the entire study area, there is no substantial difference in the simulated latent heat between the experiments with and without the lakes due to the limited lake area compared to the entire study region. However, the latent heat in the entire study area simulated by the experiment with the lakes is 65.1 W m⁻², slightly higher than 61.2 W m⁻² in the experiment without the lakes.

There is a change in the sensible heat flux over the lake area after removal of the lakes (Figure 3B). The results from the experiment with the lakes show that the sensible heat is the lowest (15.4 W m⁻²) in June among all months. After removal of the lakes, the sensible heat in June increases to 76.7 W m⁻². In addition, the sensible heat increases substantially in other months. The average sensible heat in the lake area between May and September simulated by the experiment with the lakes is 30.1 W m⁻². Without the lakes, the average sensible heat over the lake area between May and September increases by 118% in 65.7 W m⁻². From a perspective of the entire study area, there is no substantial difference in the sensible heat simulated by the experiments with and without the lakes due to the limited lake area compared to the entire study region. However, the sensible heat in the entire study area simulated by the experiment with the lakes is 67.5 W m⁻², slightly lower than that simulated by the experiment without the lakes (68.9 W m⁻²).

Overall, the lakes tend to decrease the sensible heat flux and increase the latent flux in the lake area. During the simulation period (May to September 2013), the average sum of the sensible and latent heat fluxes in the lake area simulated by the experiment with (without) the lake is 185.8 (130.4) W m⁻², indicating that the presence of lakes leads to a considerable increase in the total amounts of heat and water vapor exchange at the land–air interface. From a perspective of the entire study area, the sum of the sensible and latent heat fluxes simulated by the experiments with (without) the lakes is 132.6 (130.1) W m⁻², suggesting that there is also a slight decrease in the sum of the sensible and latent heat fluxes after removal of the lakes due to their limited area.

The height of the ABL over the area of Gyaring and Ngoring lakes and the surrounding land surfaces simulated in the control experiment is ∼500 and above 1,500 m, respectively (Figure 4A). The height of the ABL over the lakes increases to 2,000 m in the no lake experiment (Figure 4B). The presence of the lakes effectively reduces the height of the ABL over the lake area during daytime.

Figure 5 shows the distribution of the average wind field at 14:00 LST. Local dispersion wind field can be noted over Gyaring and Ngoring lakes; it flows from the center of the lakes to the surrounding land. The dispersion wind field is absent in the no lake experiment, implying that Gyaring and Ngoring lakes can affect the wind field characteristics near the surface.

The precipitation over Gyaring and Ngoring lakes simulated by the control experiment is high in the south and low in the north (Figure 5). Compared to the control experiment, the precipitation over the central lake area simulated by the no lake experiment is increased by more than 20 mm over some locations (Figure 6). The increase reaches up to more than 25% of the precipitation in this area for the same period.

Figure 7 shows prevailing westerly winds at 550 hPa over the Gyaring and Ngoring lakes with the air temperature decreasing from south to north. Compared to the control experiment, the no lake experiment produced slightly increased air temperature in most parts of the study area with a maximum increase of ∼0.15°C located in the south of Gyaring and Ngoring lakes.

Figure 8 shows the distribution of surface types included in the WRF model in the second nested domain of the study area. Grasslands account for the highest proportion (76.3%) of all surface types, followed by shrublands (11.8%). Mixed shrubland/grassland composite zones and water bodies account for 7.9% and 3.5% of all surface types, respectively.

The external water vapor sources (those outside the second nested domain) account for 63.8% of the water vapor sources of precipitation in the entire study area and are the primary water vapor sources for precipitation (Figure 9). The evapotranspiration from the grassland surfaces is the secondary water vapor source for precipitation, which accounts for 25.2% of the total water vapor sources. In addition, only 4.2% of precipitation in the Gyaring and Ngoring lakes is contributed by the evaporation from the lakes.

Precipitation in the entire study area is mainly affected by the evapotranspiration from grassland surfaces (except for the external water vapor sources), which contributes to the precipitation in almost all parts of the second nested domain area (Figure 10), while the evaporation from the lakes is the primary water vapor source for the precipitation over the lakes. Precipitation induced by evaporation from the lakes also mainly occurs within a limited area that comprises the lakes and their surrounding area. Evapotranspiration from the shrubland surfaces, grassland/shrubland composite surfaces, sparsely vegetated surfaces, and tundra surfaces contributes relatively insignificantly to precipitation.