| Deep Learning | + | Data |
MIT Parking Prediction
Parking prediction using parking reader records
Team: Tianyu Su, Zhuangyuan Fan
Supervisor: Prof. Jinhua Zhao
# Deep Learning, Travel Behavior Modeling,
Spatial Analytics, Mapping
Boston region has been challenged by long commute time and increased parking demands, and in recent years, MIT campus saw a rising issue of providing enough subsidized parking for faculty and employee. In order to maintain a sustainable campus, the school office of sustainability initiated various travel demand management (TDM) programs such as Access MIT and carpooling to encourage travel mode shiftings among MIT faculties, employee, and students. While most of these programs and initiatives are adopted across the whole institute, the 2019 MIT Commuting Survey shows that individual travel mode choices and their responses towards the same program vary significantly. This phenomenon provides us with a chance to consider mirco-targeting strategies designed for different subgroups.
In order to design an actionable “tailored” commute-related recommendation/program, it is crucial to understand the current spatial-temporal pattern of campus parking and its relationship to the regional context. In this project, we focus on predicting the number of daily gated parking coming from each census tract in Parking Year 2017-2018 (Sep. 16, 2016 - Sep. 15, 2018), as well as the overall number of daily parking of all parking permit holders. We trained a dual-stage attention-based recurrent neural network (DA-RNN) model with the campus parking history from each census tract, and exogenous variables such as weather, holidays, transit health, and use the model to predict the daily number of parking coming from each census tract. We also conduct a comparison of the baseline ARIMA model and our DA-RNN model using the same variables.