B. ANALYSIS OF NWP PREDICTION ERROR 49
3. Layer 1 Temperature 81
Systematic NWP error causing RH predictions to be too low could be due to temperature predictions that are too warm, moisture predictions that are too low, or a combination of both. Distributions of predicted and observed layer 1 temperature for each region are shown in Figures 39–41. In the coastal region, the NWP model climatology from every member is shifted several degrees warmer than the observed climatology, resulting in a clear warm bias. Seven of the members had no predictions <276 K, yet the observed climatological incidence of temperatures below this threshold is 0.2019. The same deficiency is present in the valley region, although it appears to be less severe in most of the members. The distributions of predictions in the mountain region do not show a clear warm bias. The mountain region is also unique for its bimodal distribution of observations, a feature also reflected in the prediction distributions of most members.
The verification rank histograms for the layer 1 temperature (Figure 42) show that the stochastic predictions from the entire ensemble suite also have a clear warm bias in both the coastal and valley regions. The bias in the coastal region is the most severe, with over 70% of the observation verifying below every member’s prediction. A minor warm bias is evident in the mountain region. The ensemble is shown to be underdispersive in each region.
Figure 39. Histogram of distribution of NWP layer 1 temperature predictions (blue bars), and observations (green bars) for coastal region. The first six hours of each case are
Figure 40. Same as in Figure 39, but only for the valley sites.
Figure 42. Verification rank histograms of layer 1 temperature for the coastal region (left), valley region (center), and mountain region (right). The first six hours of each
case are excluded.
Figure 43 confirms that, when all the data is included (left column), the coastal region exhibits the largest warm biases, which gradually increase during the night and reach nearly 5 K by all members just prior to sunrise. The pattern is the same with less magnitude in the valley region, with the warm biases reaching 2–3 K for most members before returning to near-neutral after sunrise. The nature of the error variances, however, is quite different between the two regions. At the coastal sites, the error variances reach nearly 20 K2 pre-sunrise, then decrease to about 5 K2 during the late morning. In contrast, error variances in the valley region are relatively low overnight, then increase by 5–15 K2 after sunrise. This pattern closely follows those of the layer 1 RH error variances in each respective region, suggesting the temperature prediction errors are at least partially responsible for the layer 1 RH errors.
To compare these results more closely in the context of diurnal temperature changes, Figure 44 shows the mean temperature change of observations (green) and predictions (blue) during the interval 7–15 h (2300-0700 LT), and again from 15–20 h (0700-1200 LT) for all cases. Although it is mean temperature changes that are shown, the line for the predictions does not start at zero but has been displaced upward above the line for the observations so that the mean bias of the predictions is also portrayed throughout the plots. The thin dashed lines represent ± 1σ of the temperature changes (not the biases) for each of the two intervals. The plots show that both regions exhibit mean observed diurnal temperature changes of several degrees, but the valley region
with observations. This is a different result than that achieved by Tardif (2007), who found delayed fog onset due specifically to inadequate cooling.
In contrast, the coastal region predictions have a total temperature range that averages <1 K across the entire post-spin up period (7–20 h), suggesting a general deficiency in the handling of boundary layer temperature forcings. The difference between the two regions is especially evident during the interval 15–20 h, when the coastal region predictions show mean warming of only 0.8° C.
Figure 43. Layer 1 temperature bias and error variance for each member for coastal (top two rows), valley (center two rows), and mountain (bottom two rows) regions. The left column shows all data, the center column includes only fog hits, and the right
Figure 44. Layer 1 mean temperature change for observations (solid green line) and predictions (solid blue line) from 7–15 h, and again from 15–20 h in the coastal region (left) and valley region (right). The line for the mean prediction change is
offset above the line for the mean observations change so that the mean bias of the predictions is also portrayed throughout the plot. The dotted lines represent ±
1σ of the temperature change within each interval.
It is proposed the difference in overnight error variances between the two regions is attributable to a more consistent nighttime boundary layer structure in the valley region, which is subject to large-scale radiative cooling and weak drainage flow on the majority of nights, as opposed to a mix of less-consistence radiative cooling and advection complicated by larger low-level temperature gradients inherent to the coastal region. While the valley region structure is seemingly more predictable for the WRF members than the coastal boundary layer, the warm bias in both regions suggests the NWP model does not fully resolve the coldest air near the surface. Perhaps the coastal region predictions are also sensitive to IC bias, which is shown to average 3–4 K warm in all members. After sunrise, the decreasing error variances in the coastal region are due to observed warming that is more consistent in timing and amplitude, whereas warming in the valley region has more day-to-day variation not resolved by the predictions.
The biggest reason for greater variation in warming rates in the valley region may be the greater tendency for fog to linger well into the late morning, with most cases absent in the predictions; at 20 h, the incidence of observed fog is 0.2338 in the valley region, and only 0.0893 in the coastal region. These post-sunrise trends are consistent
with the fog BSSs in each region, which generally increase after sunrise in the coastal region, and decrease in the valley region.
Examining the layer 1 temperature biases and error variances for fog hits (Figure 43, center column) and fog missed opportunities (Figure 43, right column) shows that the members in both the coastal and valley regions have virtually no temperature bias when fog is correctly predicted. However, predictions resulting in fog missed opportunities are characterized by a warm bias of at least 3 K at most hours in both regions. This disparity is additional evidence that the observed temperature deficiencies are linked to qc prediction deficiencies (via RH prediction deficiencies).
For most members, the layer 1 temperature biases in the coastal region are larger by <1 K larger during missed opportunities compared to the biases for all the data. This aspect of the predictions makes layer 1 temperature a good candidate for a bias correction in this region. However, the large overnight error variances are a drawback, as are the abrupt change in biases after sunrise. Even with a successful bias correction, the full impact on improving the skillfulness of the RH and qc predictions is also dependent on the nature of the water vapor predictions, which are examined in the next section.
The layer 1 temperature predictions are perhaps slightly less suitable for a bias correction in the valley region given the larger overnight biases by 0.5–1.5 K during missed opportunities compared to the biases for all the data (the differences become larger after sunrise). However, the reasonably consistent nature of the biases as a function of forecast hour, and the low error variances relative to the coastal region are positive characteristics of the predictions that might be leveraged to inform qc adjustments using methodology other than a bias correction. Whether this is the case is explored in subsequent chapters.