This paper presents an analysis of convective weather impact forecasts on traffic crossing Airspace Flow Program (AFP) boundaries FCAA05 (eastbound traffic crossing the western boundary of the Cleveland Air Route Traffic Control Center (ARTCC)) and FCAA08 (northbound traffic passing through the Washington DC ARTCC). VIL and echo top forecasts from CoSPA are translated into Weather Avoidance Field (WAF) forecasts using the en route convective Weather Avoidance Model (CWAM). Blockages on each route traversing the AFP boundary are calculated from the WAF, using the RAPT route blockage algorithm. The blockages for all routes through the AFP boundary are then aggregated into a single AFP blockage. Blockage truth' (based on actual weather) and forecasts (1, 2, 3, , 7 hours) were calculated hourly. The accuracy and volatility of blockage forecasts for the individual routes and the overall AFP were analyzed. Although this study considers only the FCAA05 and A08 AFP boundaries, the methods may be readily adapted to any arbitrary FCA boundary.
This paper also presents a method for predicting both the error and the error bounds of the AFP blockage forecasts using a Regression Tree (RT). The input (predictor) variables to the RT include the time-to-valid, the hour of the forecast issuance, the aggregate AFP blockage forecast, the spatial variance of the blockage forecast (calculated along the AFP boundary), the current AFP blockage (at the time of issue), and the current spatial variance in blockage across the AFB boundary. Each terminal node of the RT contains both the mean and standard deviation of the subset at the node, which correspond to the predicted error and error bound, respectively. The RT is compared against a simple alternative model of calculating the historical mean and standard deviation of the error for each time-to-valid. Results show that using the RT reduced the mean error bound by 11% compared to the simple model while both models were tuned so as to correctly encompass 90% of the forecasts. In other word, the RT provided the same level of correctness while providing a significantly tighter average bound.