Comparing the RWFS forecasts with and with out the FSL supplemental models (black
solid and dashed lines, respectively), the RMSE values were very similar at each lead
time for most fields. The RWFS without the FSL models is slightly better for a few lead
times for the temperature and wind variables (see Fig. 11.1). For cloud cover the FSL
models (MM5 and WRF), themselves, performed better than the other components. Case
studies elsewhere in this document demonstrate that the FSL models contribute the most
to precipitation type and timing forecasts.
Some trends are apparent in the bias plots for the season (Fig. 11.2). Positive bias (over-
forecast) is evident in the RWFS for temperature, dew point, and wind speed around
sunrise (6am-8am; 18-20 and 42-44 hr lead times). This bias quickly turns to negative
(under-forecast) during the remainder of the daylight hours (8am-6pm; 20-30 hr lead
times). The negative temperature bias is strongest during the afternoon hours (24-30 hr
lead times). For temperature and dew point, a slight, but consistent negative bias was
present during the coldest part of the night (12am-6am; 12-18 hr and 36-42 hr lead
times). A consistent, weak positive wind speed bias was also present at those same times,
and a persistent negative bias was present for wind-u and a positive bias for wind-v at all
lead times. For cloud-cover there was essentially no bias at any lead time, implying that
the system did not consistently over- or under-forecast cloud cover. However, Fig. 11.1
reveals that the cloud cover forecasts were consistently off by ~0.3, or roughly one cloud
cover category (e.g. broken instead of overcast).
Results and Recommendations: The data fusion methods and statistical
techniques utilized in the Road Weather Forecast System (RWFS) improves
the overall weather prediction skill for parameters that are measured and
available in real time. The use of multiple inputs also makes the system more
robust as it is not prone to down time with the loss of individual forecast
modules. The use of techniques and methods similar to those of the RWFS is
recommended. Dynamic weighting of the cloud cover field may be more
robust than using fixed weights based on forecaster experience.
96
Previous Page
