空中交通流量管理 (ATFM) 延誤演變預測——NM 運營 AI 工具如何改變運營.pdf

《空中交通流量管理 (ATFM) 延誤演變預測——NM 運營 AI 工具如何改變運營.pdf》由會員分享，可在線閱讀，更多相關《空中交通流量管理 (ATFM) 延誤演變預測——NM 運營 AI 工具如何改變運營.pdf（12頁珍藏版）》請在三個皮匠報告上搜索。

1、?Forecast of ATFM delay evolution FADE How an NM operational AI tool transforms operationsCamille ANORAUD EUROCONTROLHannes STAUDACHER Austrian Airlines?CONTEXTCONTEXTOperations pain points:Airspace users only know the current ATFM delay The ATFM delay assigned to a flight may change with time This

2、uncertainty has operational and economic consequencesTime before EOBTScope:Predict the evolution of ATFM delay for regulated flightsPREDICTED DELAYPROBABILITY TO DECREASEPARTNERSPARTNERSMODELMODELInputs:EFDs:city-pair,time to EOBT,airline,day of the week.Regulation(s)affecting the flight:reference l

3、ocation,reason.Training:Training dataset cleaned from obvious human actions?FADE predicts what is going to happen if the dispatcher does not act1st version-model too optimistic with high delays?Increased the weight for these cases?DEPLOYMENTDEPLOYMENT?Model deployed on NM Cloud platform(Cloudera)Mon

4、itoring of performance by NMDaily monitoring of performanceAutomatic retraining once the performance reaches the threshold DEPLOYMENTDEPLOYMENT+80 airspace usersNMP/NMUI Flight May 2023APIAPI for AI application from NM?SATISFACTION SURVEYSATISFACTION SURVEY Satisfaction survey launched in October 20

5、24 and sent to the PoC 41 AUs answered out of 65 PoC 63%response rate?Some improvements to make:Reliability and accuracy Further information on FADE:Reports on FADE performance Limitations of the models:adverse weather,ATC-industrial actions Usage of the models:timeframe,probability thresholds,model

6、s complementarityVery positive feedback:Useful tool with great potential Well-appreciated and regularly used Big improvement compared to before(no information on delay evolution)Very good predictions,generally accurate during Summer operations?SATISFACTIOSATISFACTION N SURVEYSURVEY?NEXNEXT T STEPSST

7、EPSImprove accuracy Review of the training process of the model:time binning strategy,pinball loss instead of MAE Consideration of new metrics for performance monitoringImprove communication on FADE Limitations of the model Usage of the models:probability thresholds,models complementarity Performanc

8、e reporting directly in NMUI Flight?USERSUSERS JOURNEYJOURNEYInternalBuilding up confidenceInternal becoming the prio list for our dispatchSupports tactical planning and steering Prediction generates expectationPrioritizing regulated flights Shifting resources to where it mattersSlot Coordinator?Thank you!?