9.1 A Response Surface Model……………………………………………………9.3 Models With Different Numbers of Weather Factors …………………………LAXReport.pdf1. Introduction2. Methodology Overview3. Computing the Daily Flight Time Index for LAX4. Weather Normalization VariablesFigure 5. LAX Airport ConfigurationFigure 6. West Flow at LAX2225.pdf5. Demand Normalization Variables3136.pdf6. Origin Airport Delay Normalization7. Regression Modeling of DFTI at LAX3867.pdf8. Regression Modeling of DFTI Components9. Alternative Model Specifications9.1 A Response Surface Model9.2 A Non-linear Model9.3 Models With Different Numbers of Weather Factors9.4 Models With Non-Parametric Weather EffectsFigure 16. Illustration of Cluster Analysis for LAX10. OutliersTable 18. Summary for Outliers11. Conclusionsabstract.pdfAbstractInstitute of Transportation StudiesUniversity of California at BerkeleyJune 2001ISSN 0192 4095RESEARCH REPORTUCB-ITS-RR-2001-5Delay and Flight Time Normalization Procedures for MajorAirports: LAX Case StudyMark HansenTatjana BolicNATIONAL CENTER OF EXCELLENCE FORNEXTORAVIATION OPERATIONS RESEARCHAcknowledgmentsThis report documents research supported by the National Center of Excellence inAviation Operations Research (NEXTOR) under FAA Contract DTFA03-97-D-004 andCooperative Agreement 01-C-UCB-02. The authors would like to thank Dave Knorr,Rich Gutterud, and Ed Meyer of the FAA and Joe Post from the Center for NavalAnalyses for their support and encouragement in this project. The opinions expressedherein are those of the authors and may not reflect those of the persons or organizationsjust cited. The authors are also solely responsible for any errors or omissions in thisreport.Abstract This report presents methodologies for normalizing performance of the National Airspace System (NAS). The purpose of the study is to develop the capability of isolating the performance of NAS enhancements, such as those being made under the Free Flight Phase I program. It is often not possible to observe the effect of such enhancements directly, because of the confounding influences of weather, demand, and conditions elsewhere in the system. The analysis presented here shows how linear and non-linear regression models can be used to statistically remove a large proportion of these confounding effects, increasing the possibility that the effects associated with the enhancement will be detectable. The particular focus of this study is on arriving flights at Los Angeles International airport, where two FFP1 tools, Traffic Movement Advisor (TMA) and Passive Final Approach Spacing Tool (PFAST), are being deployed. We develop a metric that captures the daily variation in flight times (including departure delay and gate-to-gate time) for LAX arrivals. This metric, which we term the Daily Flight Time Index (DFTI), is a weighted average where the weights reflect the proportions of flights coming from different destinations over the analysis period. We then analyze the day-to-day variation in DFTI, relating it to weather, demand, and average delays at origin airports. Our data set extends over 41 months from January 1997 through May of 2000. Our approach was to develop a “baseline” model and then compare it with a variety of others. The baseline model contains 9 weather factors (scores from which are generated from applying principal component analysis to 32 underlying weather variables), 2 demand factors, and an origin airport delay variable, in a simple linear form. It explains about 75% of day-to-day variation in DFTI. Origin airport congestion is the most important source of variation, followed by several weather factors relating to temperature, visibility, and wind. Demand is the least important source of variation in DFTI over the time period analyzed. Most of the effects observed are intuitively reasonable: for example, we find that the DFTI decreases with visibility. Some are more mysterious--for example, DFTI is found to decrease with temperature at LAX. Several other models are estimated and compared to the baseline model. A response surface model that includes quadratic and interaction terms as well as linear ones offers some improvement in fit (adjusted R2 of 0.82 as compared to 0.74) albeit with a vastly increased number of coefficients. A non-linear model also performs somewhat better. Model performance is quite insensitive to the number of weather factors used. However, models that capture weather by categorizing days rather than employing quantitative weather factors perform somewhat less well. Finally, TRACON logs from outlier observations were inspected to find reasons that the model predictions were inaccurate for these days. Generally, it was possible to discern explanations on days when the model under-predicted the DFTI. These included facility outages, overly stringent ground delay programs, east flow operation, and, in one case, the closure of half the airport due to a visit by Air Force 1.i TABLE OF CONTENTS: 1. Introduction………………………………………………………………………... 1 2. Methodology Overview…………………………………………………………… 3 3. Computing the Daily Flight Time Index for LAX………………………………… 5 4. Weather Normalization Variables…………………………………………………. 11 4.1 Qualitative Discussion………………………………………………………... 11 4.2 Normalization Procedure…………………………………………………….. 17 5. Demand Normalization Variables…………………………………………………. 23 6. Origin Airport Delay Normalization………………………………………………. 34 7. Regression Modeling of DFTI at LAX……………………………………………. 36 8. Regression Modeling of DFTI Components………………………………………. 42 9. Alternative Model Specifications………………………………………………….. 46 9.1 A Response Surface Model…………………………………………………… 47 9.2 A Non-Linear Model………………………………………………………… 50 9.3 Models With Different Numbers of Weather Factors ………………………… 52 9.4 Models With Non-Parametric Weather Effects………………………………