Tuesday, 24 January 2012: 11:00 AM
Prediction and Correction of Convective Weather Impact Forecast Errors Using Regression Tree Ensembles
Room 242 (New Orleans Convention Center )
Extended (0-8 hours or more) convective weather forecasts, such as CCFP, LAMP, and CoSPA, have been developed to support strategic planning of air traffic flows that avoid serious weather impacts and maintain an efficient balance between demand and capacity. Strategic planners use Flow Constrained Areas (FCAs), lines or polygons to identify regions of airspace that may be impacted by severe weather. Traffic managers develop plans to limit traffic flows through FCAs based on their best estimate of future capacity impacts, given the convective weather forecasts available to them. A model has been presented to translate CoSPA forecasts of Vertically Integrated Liquid (VIL) and echo top heights into objective predictions of weather impacts and forecast uncertainty for the capacity of flows through an FCA based on the blockage of individual routes crossing the FCA boundary . This model has also been proposed as an operationally relevant metric for forecast evaluation. This paper describes an extension to that model that applies a quantile regression random forest to time-lagged ensembles of capacity impact forecasts, based on CoSPA weather forecasts, to eliminate bias, reduce error, and predict error bounds of the capacity impact predictions. Regression model inputs were taken from time lagged ensembles of capacity impact forecasts for the full FCA and individual air traffic routes crossing the FCA. Models were trained on a dataset that included several thousand capacity forecasts from 2010 and 2011 across two commonly used FCAs: FCAA05, which controls eastbound flows into Boston, New York, and Potomac, and FCAA08, which controls northbound flows into Potomac and New York. The random forest approach presented here has several desirable properties. It addresses the problem inherent in the use of time-lagged ensembles to assess forecast uncertainty and reduce forecast error – that ensemble members are highly correlated and unlikely to span a meaningful range of possible outcomes – by predicting outcomes based on the correlation of the current ensemble to a large historical database of ensemble inputs and observed outcomes. Error bounds for a given forecast can be determined by examining the trees generated by the forecast algorithm. Forecast and error bound prediction models are easily trained, and the presented algorithm can be easily adapted to any set of input ensemble forecasts from which route blockage can be calculated. Capacity impact forecast error and bias reduction, and the accuracy of forecast error bounds are compared for different random forest models. The sensitivity of model performance to the underlying airspace structure and the relative skill of different model inputs in predicting capacity impacts are examined. Finally, case studies illustrating model results in different weather scenarios are presented. References 1. Lin, Yi-Hsin, Joshua Sulkin, Rich DeLaura, “Prediction of Weather Impacts on Air Traffic through Flow Constrained Areas”, American Meteorological Society Special Symposium on Weather and Air Traffic, Seattle, WA, 2011.
This work was sponsored by the Federal Aviation Administration under Air Force Contract No. FA8721-05-C-0002. Opinions, interpretations, conclusions, and recommendations are those of the authors and are not necessarily endorsed by the United States Government.