A DYNAMIC BAYESIAN BELIEF NETWORK APPROACH FOR MODELING THE ATM - - PowerPoint PPT Presentation
A DYNAMIC BAYESIAN BELIEF NETWORK APPROACH FOR MODELING THE ATM NETWORK DELAYS Yi it Bekir Kaya Data Science Researcher at CAL, Aeronautics Graduate Student Prof. Gkhan nalhan Director of CAL Istanbul Technical University Controls
A DYNAMIC BAYESIAN BELIEF NETWORK APPROACH FOR MODELING THE ATM NETWORK DELAYS
Yiğit Bekir Kaya Data Science Researcher at CAL, Aeronautics Graduate Student
Director of CAL Istanbul Technical University Controls and Avionics Laboratory Aeronautics Research Center
¨ Modeling of ATM Network Delays ¨ Identifying Patterns and Best
Practices for Resilience against System Upsets (Resilience2050.eu)
¨ Creating a stochastic model that can be used as the basis for
¤ Dynamic Slot Management (SecureDataCloud SESAR WP-E) ¤ A-Collaborative Decision Making
¨ Giovanni Bisignani (CEO of IATA) claims:
“Shaving one minute off each commercial flight would save 5.0 million tons of CO2 emissions and $3.8 billion in fuel costs each year”
¨ For airlines having about 2% of market share (e.g. THY), the saving
is $76 Million per year
¨ Weather
¤ Capacity Decrease
n Runway Change n Change in movements per hour ¨ ATC Capacity ¨ Aerodrome Capacity ¨ Environmental Issues
¤ Volcano eruption
¨ Special Events
¤ Airspace closure
n Military
¤ Airline strikes
¨ ATC Staffing ¨ Accident/Incident ¨ Airspace Management
sector in any demanded level and may include set of aerodromes (airports)
the same sector is represented with a loop
are represented with sources and sinks in each sector block. In this regard, whole sector block can be deemed as delay (and traffic) generator/consumer which consists of mini generators.
(FTFM), generated delays due to sector capacity/restriction/ traffic overflow is obtained for each Flight
FIR segments
at aerodromes (or at TMAs)
¨ ρ is utilization rate ¨ If ρ < 1 the system is at steady
state and
¤ proportional to
¨ Otherwise the system is
chaotic
¨ Initial Delay Perception
¤ AOBT – EOBT (delay based on estimation; EOBT = IOBT 96% in data)
¨ Strict Definition Delay (Pushback, Gate-out delay)
¤ AOBT – SOBT
¨ Passenger Perceived Delay
¤ ATOT – STOT
¨ Taxi-out Delay
¤ (ATOT - AOBT) – (STOT - SOBT)
¨ Taxi to TMA Exit Delay (ADTET = Actual Departure TMA Exit Time)
¤ (ADTET – ATOT) – (SDTET – STOT)
¨ Departure Delay
¤ Pushback Delay + Taxi Delay + Taxi to TMA Delay ¤ ADTET – SDTET
¨ En-route Delay (AATET = Actual Arrival TMA Enter Time)
¤ (AATET – ADTET) – (SATET – SDTET)
¨ TMA Entry to Taxi Delay
¤ (ATOA – AATET) – (STOA – SATET)
¨ Taxi-in Delay
¤ (AIBT – ATOA) – (SIBT – STOA)
¨ Gate-in Delay
¤ AIBT – SIBT
¨ Arrival Delay
¤ TMA Entry to Taxi Delay + Taxi-in Delay + Gate-in Delay ¤ 2*(AIBT – SIBT) – (AATET – SATET)
¨ Passenger perception (*)
¤ STOT as departure time ¤ SIBT as arrival time
¨ The ALLFT+ data set is managed by the PRISME (Pan-European
Repository of Information Supporting the Management of European Air Traffic Management Master Plan)
¨ Every entry is a single flight information
¤ Flight Plan ¤ Tactical Flight Model
n FTFM n RTFM n CTFM
¤ Routes
n CPF-GEN n CPF-REF
¨ AOBT and EOBT as is ¨ IOBT = STOT ¨ SOBT = STOT – nominal time (e.g. 15 min ~ airport) ¨ SFP (Planned Flight Profile, FTFM), AFP (Actual Flight Profile, CPF/
CTFM)
¨ ATOT= first radar point (AFP[0].entryTime)
¤ 0 stands for first entry and -1 stands for last entry (Circular array
notation)
¨ ADTET = AFP[0].exitTime, SDTET = SFP[0].exitTime ¨ AATET = AFP[-1].entryTime, SATET = SFP[-1].entryTime ¨ ATOA = AFP[-1].exitTime, STOA = SFP[-1].exitTime ¨ AIBT
, SIBT are not in ALLFT+ data
¤ Taking (AIBT – SIBT) as nominal (gate-in)
¨ Linear/Nonlinear Regression ¨ Graphical Models
¤ (Dynamic) Bayesian Belief Network ¤ Hidden Markov Models ¤ Kalman Filter
¨ Time Series Model
¤ SARIMA, GARCH
¨ Nonparametric Methods
¤ Nonparametric Density Estimation
n Kernel estimator, Histogram, k-NN
¤ Smoothing models
n Mean, kernel, running line, moving median, smoothing splines
¤ Multilayer Perceptrons
¨ Decision Trees
¤ Random Forest
¨ P(G,S,R)=
P(G|S,R)P(S|R)P(R)
Belief Propagation:
¨ P(X|E)=
αP(X|E+)P(E−|X) = απ(X)λ(X)
¨ Big Picture Approach
¤ No assumptions about inner models of airports ¤ OD pairs are analyzed independently (Eulerian Approach)
¨ Pure Bayesian Model
¤ No assumption about mathematical structure of delay propagation ¤ Observation (data evidence) based probabilistic model
¨ Time Behavior
¤ There is a stochastic relationship between lags
¨ Departure delay prediction benefits from Belief Propagation more
than other phases
¨ There is a ±22.5 min margin of error from Departure Delay
Prediction for 95% confidence interval
¨ More data samples are needed for accuracy increase ¨ More information should be provided to the system in order to
model underlying system
¨ Weather and Capacity Data should be aggregated along with Delay
Data
SARIMA, GARCH
¨ Two timing approach
¤ Seasonal periodicity (s) ¤ Hourly periodicity (t)
¨ Delay = f(s, t) = Φ(s) + Θ(t) + w ¨ fbar(s) = daily mean of f(s, t) ¨ Φ(s) = SARIMA(fbar(s)) + WeatherModel(fbar(s)) ¨ f’(t) = hourly mean of {f(s, t) - Φ(s)} (Making levels even) ¨ Θ(t) = SARIMA(f’(t)) + QueueModel(f’(t)) + WeatherModel(f’(t)) ¨ w = f(s, t) - Φ(s) + Θ(t) ¨ w ~ N(0, σ)
¨ SARIMA
¤ AR - Auto regressive ¤ MA - Moving Average ¤ I - Integrated ¤ S – Seasonal
¨ Conditions
¤ TS should be linear ¤ TS should be stationary ¤ TS should not have any trends (detrending) ¤ TS should be significantly different than white noise ¤ Residuals should be white noise
¨ Condition Analysis
¤ Non-linearity: White Test ¤ Stationary/Explosive: Dicky-Fuller Test ¤ White Noise: Box-Jung Test ¤ Seasonality: Auto Correlation Function ¤ Cross Correlation
¨ AR(p)
¤ xt − µ = φ1(xt−1 − µ) + φ2(xt−2 − µ) + ··· + φp(xt−p − µ) + wt,
¨ MA(q)
¤ xt = wt + θ1wt−1 + θ2wt−2 + ··· + θqwt−q,
¨ ARMA(p, q)
¤ xt = α + φ1xt−1 + ··· + φpxt−p + wt + θ1wt−1 + ··· + θqwt−q
¨ A Sample Equation for ARMA(2,2) Model:
¤ xt = .4xt−1 + .45xt−2 + wt + wt−1 + .25wt−2 ¤ Where xt denotes dependent time series and wt denotes white noise
time series
¤ wt~N(0,σw
2) ¨ ARIMA(0; 0; 0)x(0; 0; 1)12 ¨ GARCH: Similar to ARIMA
¤ Generalized Auto Regressive Conditional Heteroskedasticity ¤ Heteroskedasticity: No constant variance assumption ¤ The variance can be estimated ¤ Yt = f(X1,t;… ; Xp,t) +σ(X1,t; … ; Xp,t) wt;
¨ Using only Bayesian Network causes higher margin of error than
SARIMA model for the same confidence interval
¨ However, Bayesian Network provides a probability distribution
rather than only mean and standard error.
¨ Bayesian can process missing values ¨ Belief propagation might decrease the variability of the result ¨ Random Forest provides importance information ¨ Non parametric methods such as Multilayer Perceptrons are highly
dependent on current data and does not provide a parametric inference
¨ Non parametric methods are very sensitive to initial state ¨ Non parametric methods can “over-fit” data ¨ Non parametric benefit from online learning
¨ Unstructured
¤ Data has no structural information in it
n No labels for features (or fields) n No hierarchy between fields ¨ Text formatted files
¤ BigData is stored in plain text files ¤ Fields are separated by symbols or white space characters
¨ Very large at size
¤ Daily information is about several Gigabytes for flight information ¤ A simple aggregated database can reach to Terabytes
¨ Data generation has very high speed
¤ Collecting data is far more faster than analyzing them ¤ There is not enough bandwidth to transfer data to outside servers for
processing BigData in reasonable time
¨ Diverse source of information collection
¤ For aviation every stakeholder have their own interest of collecting
data
¤ Aligning and synchronizing data can be cumbersome
ALL_FT+ Data set includes eight different Airspace profile; Tactical flight Models
¨ FTFM - Filed Tactical Flight Model; The FTFM is the “initial” profile
as it reflects the status of the demand before activation of the regulation plan. It is computed with the latest flight plan version, sent by each AO to the CFMU/IFPS
¨ RTFM - Regulated Tactical Flight Model; The RTFM is the
“regulated” profile as it reflects the status of the demand after activation of the regulation plan. It is computed with the latest ATFM slot (CTOT) issued to the AO, by the ground regulation system
¨ CTFM - Current Tactical Flight Model; The CTFM is the “actual”
profile as it integrates the actual entry time of the flights in the regulated TV. It is computed with the Radar Data sent by ACCs to CFMU/ETFMS ref
¨ CPG_GEN - Profiles generated by the CFMU path generation tool
¤ SCR - Shortest Constrained Route ¤ SRR - Shortest RAD restriction applied Route ¤ SUR - Shortest Unconstrained Route ¤ DCT - Direct route
¨ CPF - Correlated Position reports for a Flight; CPRs (Correlated
Position Reports) which are surveillance data collected from the ACCs.
¨ BMT converts unstructured text file to managed and efficient
database
¨ BMT allows BigData to be consumed efficiently by applications
processing BigData provided
¨ BMT can process any BigData which is in text format separated by
symbols
¤ Other extensions can be built for other possibilities
¨ Stores data in MongoDB
¤ Schemaless design ¤ Flexible, distributed ¤ Highly efficient and secure
¨ Reduces the size of data stored
¤ Up to 5-6 times (with extensions the size of data can be further
reduced up to 60-140 times with some trade-off issues)
¤ Reducing size and distributed structure of MongoDB makes data
processing ultra faster depending on the scale of the data
¨ BMT eliminates client application’s parsing overhead at every run
¤ Structured information is stored in efficient-to-process data centers
¨ By utilizing Python rather than MATLAB the parsing time of a single
day of ALLFT+ data is dropped from 5 hours to 14 minutes at the same specs
¤ It can be further reduced by distributing database across shards and
utilizing supercomputers
¨ Calculating and monitoring CO and other greenhouse gases in
Istanbul is an ongoing project implemented jointly with ITU Eurasia Institute of Earth Sciences
¤ Emissions of various gases are calculated efficiently by utilizing flight
mode and other information processed by BMT
¨ Predicting delay propagation in an airport by combining various
machine learning techniques based on different approaches
¤ Machine learning algorithms are implemented for parallel and
distributed computation utilizing BMT