Insights from machine learning to vehicle emissions
Energy Futures Lab hosts a seminar from Ms Clemence Le Cornec of the Department of Civil and Environmental Engineering on the use of machine learning in modelling vehicle emissions.
Abstract
Despite the introduction of stringent emissions and air quality regulations, NO2 concentrations are exceeding harmful levels in a large number of cities across Europe. Discrepancies between standardised type-approval and real-world driving emissions led to the introduction of the Real Driving Emission (RDE) regulation in September 2017.
The polluter pays principle suggests that drivers entering polluted areas should be charged based on their actual emissions rather than on a per entry basis that does not account for the distance driven, influence of individual driving behaviour or the variability in emissions performance of different vehicles (e.g. Euro Standard).
In this context, existing emission models are not adequate: physics-based models are computationally expensive and often require intensive calibration, while average-speed models (e.g. COPERT) are not able to accurately capture highly non-linear behaviours associated with NOx emissions.
Consequently, there is a need to develop new emission models, able to run in real-time at a low computational cost while providing the level of accuracy required to implement a polluter-tax system. With the increase in computational power and availability of real-world driving data, Artificial Neural Networks (ANNs) present an opportunity to address these issues.
While simple ANNs architectures have been used to predict NOx emissions for selected vehicles and specific conditions, this study uses a large dataset of diesel vehicles (70 passengers’ vehicles including 36 Euro 5 and 34 Euro 6) tested in real-world conditions on the same route in Greater London to develop a real-time Long-Short Term Memory (LSTM) NOx emission models.
Vehicle driving parameters (instantaneous speed and acceleration) used as inputs to the model were obtained from an embedded GPS while emissions were measured with a Portable Emission Measurement System (PEMS). As differences between vehicles (e.g. due to different manufacturer emissions control strategies) are greater than differences due to operating conditions, a general model is not able to predict NOx emissions at the accuracy required. Vehicle-specific models (one model per vehicle type), on the other hand, present the desired accuracy but would require every vehicle to be tested in real-world driving conditions, which is not realistic at a city scale.
A clustering analysis was therefore conducted to identify and group together vehicles with similar emitting behaviours. Dynamic Time Warping (DTW), a powerful method to compare time series that might vary in speed. A k-means++ algorithm was then run and the Calinsky-Harabastz index was used in combination with the Davies-Bouldin index to evaluate the optimal number of clusters. A specific model for each cluster was then developed.
ANNs models provide a powerful tool to develop new emission models. The model demonstrates an absolute mean error of 0.05 ± 0.06 g/km (median 0.03 g/km) and a mean relative error of 19.1 ± 26.7% (median 10.7%) (c.f. Euro 6 diesel standard for NOx is 0.08 g/km). At the cost of a slight decrease in performance, the clustering analysis successfully reduces the number of models needed, even though no simple variable (e.g. manufacturer, aftertreatment) is able to explain the clusters. Clustered models present an absolute mean error of 0.06 ±0.08 g/km (median 0.04 g/km) and a mean relative error of 22.7 ± 24.6 % (median 15.1%). A comparison with existing instantaneous emission models will also be performed.
Biography
After a Bachelor (2014) and a Master (2017) degree in Environmental Engineering at the Ecole Polytechnique Fédérale de Lausanne (Swiss federal institute of technology, Lausanne), Clémence started a PhD at Imperial College London focusing on instantaneous vehicle emissions modelling and air quality related issues.
Venue
The talk will be held in Room 611 of Electrical and Electronic Engineering (building 16 On the campus map). The room is known as the Gabor Suite.
If you are entering the building from Dalby Court/through the building’s main entrance take the lift to the sixth floor, turn right through the double doors and it is near the end on your left hand side.