Spatio-Temporal Mobility Demand Forecaster (ST-MDF)

Urban mobility is key for the development of any city, particularly in the case of smart cities. Being able to predict when and where a certain number of people will have the need to start a trip allows for the optimization of any given mobility service, such as taxi, bicycle or for-hire vehicles (FHV). This repository collects source code and information about a Deep Learning model named ST-MDF for mobility demand forecasting with heterogeneous mobility services. This forecasting problem can be graphically posed as follows:

The main features of the developed ST-MDF model are that:

• it works with mobility demand mesh-grids, which provides high spatio-temporal resolution (500x500m grid with 15min intervals),
• the prediction takes into account n_x previous timesteps, and forecasts demand for n_y future intervals,
• it is robust, i.e., it can recover from sensor failure and continue producing reliable predictions,
• it is flexible, i.e., it can work with a variable number of input and output sensors.

The contents of the repository are described in the following points:

• graphs: Relevant graphs about the problem, and several maps of the city chosen as the case of study: Chicago.
• notebooks: Jupyter notebooks that present how the data was analyzed and parsed. The obtained datasets are available at: https://doi.org/10.5281/zenodo.5166838
• results: Summary that includes metadata and metrics of the trained models and baselines. In the column names, MMWC is the Mean across taxi zones/bike racks of the MAE Weighted of trip Counts, per forecast horizon. Similarly, MRWC is the Mean across taxi zones/bike racks of the RMSE Weighted of trip Counts, per forecast horizon. Also, a learning rate of 0 means that the default value for that Keras optimizer was employed.
• src: It includes the following scripts:
  • deep_playground.py: Interactive script to train and test deep learning models.
  • launcher.py: Script that launches the training and testing of models iteratively.
  • learner.py: Main script that works together with modelUtils.py, trainUtils.py and plotUtils.py to manage the training process, while storing the relevant model metadata.
  • modelUtils.py: It defines a class that contains the hardcoded models, alongside the relevant model (e.g. n_x and n_y) and training (e.g. loss function, learning rate, optimizer, etc.) parameters.
  • plotUtils.py: It defines a class for plotting tasks.
  • test_gpu.py: Script to test whether the GPU/s is/are available.
  • trainUtils.py: It defines a data generator class for the training and testing of the models, and a class for keeping the relevant information of the training history.