Dr. Lama Alfaseeh, postdoctoral researcher at Ryerson University, presented “Prediction of Greenhouse Gas Emission in Downtown Toronto Using Deep Sequence Learning” on January 15, 2020 for the UT-ITE seminar series.
Transportation systems contribute the largest amounts of greenhouse gas (GHG) emissions in the US and Canada. GHG emissions, which include carbon dioxide, methane, and nitrous oxide, are more harmful to the environment than carbon dioxide alone. Predictive models for routing vehicles proactively have potential in tackling this problem. Alfaseeh lists her objectives:
- Developing a deep learning framework based on Long Short-Term Memory (LSTM) to predict GHG emission rates in a highly congested urban network using microscopic data points
- Analyzing predictor importance, i.e. how much each variable matters
- Comparing the deep learning model with common emission predictive models: autoregressive integrated moving average (ARIMA) and clustering
- Demonstrating the impact of tuning in the LSTM model
For creating the models, ARIMA uses combinations of AR, I, MA; Clustering uses random assignments; and LSTM, a recurrent neural network, uses combinations of predictors, sequences, and hyperparameters like number of hidden layers. The models were tested on data from downtown Toronto during the morning peak, 7:45 a.m. – 8:00 a.m.
In analyzing the data, the top variables highly correlated with future GHG emission rate (ER) included speed, density, current GHG ER, and in-links speed, which were used as predictors in the models. Root mean square error (RMSE) was used to measure model performance, with lower values indicating better performance.
ARIMA had the worst performance followed by clustering. Disadvantages of ARIMA were scaling and non-linearity consideration between variables, while clustering’s drawback was considering GHG ER as a discrete variable. LSTM overcomes these limitations because it is scalable unlike ARIMA and it considers GHG ER as a continuous variable unlike clustering.
The LSTM network layer when systemically tuned using Bayesian optimization outperformed the best clustering and ARIMA models. The RMSE improved by 37% and 3% compared to best ARIMA and clustering models, respectively. Using systematic tuning over manual tuning improved RMSE by 17%; this increase in performance from systematic tuning was amplified when the number of hidden layers increased from 1 to 2, improving the RMSE by 26%.
Using LSTM deep learning outperformed commonly used emission-predictive models. This increase in performance can be valuable for eco-routing on large-scale road networks to reduce the adverse impact of global warming.
Dr. Lama Alfaseeh earned her Bachelor’s Degree in Civil Engineering in 2006 and a Masters Degree in Construction Project Management in 2011 from Damascus University. She started her PhD in 2016, joined the Laboratory of Innovations in Transportation (LiTrans) at Ryerson University in 2017, and defended her dissertation in 2020. Lama was supervised by Dr. Bilal Farooq and her research investigated the impact of employing intelligent vehicles in a distributed routing environment. Lama utilized intelligent transportation systems (ITS) to mitigate the undesired effect on the environment and health. She started a postdoc position in November 2020 where she has been developing predictive models for climatic variables to help structural engineers proactively design concrete structures while considering the impact of climate change.
Presented by University of Toronto ITE Student Chapter.