Artificial intelligence in the post-deep learning era
Session 38 Xiaoliang Ma
1. Toward Integrated Traffic and Emission Impact Simulation: A Microscopic Approach Xiaoliang Ma, Ph.D. Centre for Traffic Research, KTH, Sweden Zhen Huang, PhD candidate, Centre for Traffic Research, KTH, Sweden
2. Introduction The skyrocketing of motor vehicle numbers leads to serious challenge to urban environment; Different concerns of traffic emissions: Highly dependent on fossil fuels Green house gases (GHG) e.g. CO2 Local air quality: CO, HC, NOx, PM etc.
3. Policy and Measure Long term: revolution on vehicle engine and alternative fuel technologies; Middle to short terms: Management at the network level e.g. congestion charging Local traffic measures: signal, speed limit, traffic calming… New Intelligent Transportation Systems
7. Research questions (1) Assessing traffic emissions (microscopic). Lack of emission calculation software tool: well integrated with traffic models and can estimate emission at different resolutions ; Difficulties to analyze and visualize traffic emissions and air quality impacts in time and space Traffic Simulation Output Emission Output Traffic Simulation Model Emission Model
8. Research questions (2) Microscopic emission models becomes more and more important for traffic operation and management Microscopic emission calibration requires lots of data: complex and expensive to measure Research question: How to calibrate microscopic emission models when the emission measurements are not fully available?
9. Research objectives Develop a systematic approach to integrate traffic simulation models and emission models. Develop a “calibration” method for microscopic emission models based on existing aggregate measures Evaluation and optimizing traffic measures with respect to traffic emissions.
11. Distributed implementation (Service) Better Performance Both traffic simulation and emission calculation are computational expensive processes Decentralize these into different CPU processes or threads Systematic Design Concurrency Reliability Independency Plenty of implementation CORBA Web Service (XML) Protocol buffers
14. Numerical Implementation Microscopic Emission Model VT-Micro is calibrated using aggregate estimation of ARTEMIS as reference. Newton-reflective, Finite Difference Method (FDM), Simultaneous Perturbation Stochastic Approximation (SPSA) are adopted as optimization algorithms. MAPE and MSE were used to validate model performance.
18. Various travel demands were evaluated by the calibrated microscopic emission model to assess the environmental impactsunder different congestion conditions.
28. A numerical approach is proposed for calibration of microscopic emission models based on aggregate emission measures;
29.
30. Parameters The CMEM model is comprised of six modules: 1) engine power demand; 2) engine speed; 3) fuel/air ratio; 4) engine states; 5) catalyst pass fraction. Calibration parameters Fuel rate (FR) ε1, K0,Pscale Engine-out emission (EO) αHC,αCO,αNO1 bHC,bCO,bNO Tailpipe emission (TP)
31. Data collection second-by-second tailpipe emission measurements were collected using OBS-2200, a portable emission measurement system (PEMS); 28 recruited in-use light-duty vehicles were divided into two CMEM categories:
35. VT-Micro model To compare the performance with the CMEM model, the VT-Micro model, a regression model, is also calibrated using the same datasets. Where: MOEe (g/s) is the tailpipe emission rate; i and j are powers for v and a, respectively; Lei,j and Mei,j are the regression model coefficients; v (km/h) and a (km/h/s) are vehicle speed and acceleration.
36. Model validation The Mean Absolute Percentage Error (MAPE) and the Root Mean Square Error (RMSE) were used to validate model performance. where: ykis measurement (g/s); is model prediction (g/s); N is the size of composite data samples.
45. Evaluation of signal control N=10; for C = Cmin : 5 : Cmax for L1= αmin·C : 1 : αmax·C for L2= βmin· L : 1 : βmax·L L3=C- L2- L1 Set_Vissim_SignalInput; Run_Vissim(N); Produce_Statistics; end end end Naïve grid search Minimize number of stops Minimize total delays
46. Conclusion A parameter tuning approach has been developed for the CMEM model implemented as a 3-party software; The calibrated CMEM model shows improved performance than that with default parameters according to the calibration and validation results; It is shown that by combining traffic and emission models we are able of estimating dynamic emissions and moreover optimize traffic control measures with environmental concerns.
47. Future work The parameter tuning process will be extended to other categories, especially light-duty trucks (LDT), heavy-duty vehicles (HDV), and other high emitting vehicles. Emission measurements under cold start conditions should be collected for estimating cold-start parameters in CMEM. Application of the model on improvements of traffic planning and management with respect to reduction of traffic pollution in Chinese cities.
Transportation consumes 1/3 energy supplyTrasnportation is one essential factor for climate change and local air quality issues
Microscalevehicle emission models (e.g. CMEM, VT-Micro) have been developed for estimatingvehicle exhaust emissions in the local network in a high resolution;Microscopic traffic simulation model (e.g. VISSIM) can capture traffic characteristics with a high fidelity;Combination of vehicle emission models and microscopic traffic simulation models has been adopted for assessing traffic environment;
problem
No relation between these two research questions
Characteristics:Parallel and distributed system Standardized data format for communication Online emission estimation in different resolutions Easy extension
A series of experiments are conducted in Sweden to show the on-road emission versus emission estimations of three major road transport emission models in Europe, i.e. COPERT 4 model, the Handbook of Emission Factors (HBEFA 2.1) [25] and ARTEMIS. The measurements are carried out by means of remote sensing [26], and the comparison is made on three different sites for gasoline passenger cars. It was observed that there is normally a good agreement between ARTEMIS modeled and measured emission, although there are occasionally also major discrepancies
ε1 is the maximum drivetrain efficiency;K0 determines engine friction factor during engine idling;Pscaleis a power threshold dimensionless scaling factor; αHC, αCO and αNO1 are EO index coefficients for HC, CO and NOx respectively; bHC, bCO, and bNO are catalyst efficiency coefficients.
Data from ten LDV4 vehicles and eight LDV6 vehicles are randomly chosen to build up composite calibration datasets for the corresponding vehicle classes
In stage I, the fuel-rate module is calibrated upon fuel consumption measurement by tuning three critical parameters: ε1, K0, and Pscale to minimize MSE between modeled fuel rate and measurement; In stage II, the engine-out emission module and the catalyst pass fraction module are calibrated by simultaneous tuning of three sets of model parameters, αHC and bHC, αCO and bCO, αNO1 and bNO respectively for HC, CO, and NOx emissions to minimize MSE between emission model outputs and measurements Both numerical grid search and nonlinear simplex algorithm are used in both stage of the calibration procedure; The calibration procedure has been implemented in the MATLAB environment;
The model with modified fuel-related parameters can significantly reduce the prediction errors
Although the MAPE values are not satisfactory, the prediction performance of the tuned CMEM model is significantly improved because it always shows obviously smaller MAPE and RMSE values. The predictions performance of the VT-Micro model fall between the default CMEM model and the tuned CMEM model because it produces medium MAPE and RMSE values in most cases.
The CMEM model with default coefficients cannot really capture the trends of real-world emissions; The fine-tuned CMEM model show improved performance in capturing trends of real-world emissions, but it tends to overestimates emissions during low emission profiles In most cases, the VT-Micro model shows similar instantaneous estimation performance with the default CMEM model.
The VISSIM model is calibrated for basic driving behavior parameters, including the desired speed distribution and desired acceleration distribution using traffic measurement. Vehicle composition is composed of LDV4 vehicle (50%), LDV6 vehicle (40%), truck (8%), and bus (2%). The traffic simulation model has been run with replications until the statistical measures in terms of traffic flow rates are stabilized. The number of simulation replications is determined according to the sequential procedure.