With the rise of digitalized subway cards, an huge amount of data is created and recents attempts to eploit this data have proven quite successfull. The goal of this project was two fold : first obtain a deep representation of the different subway stations of the networks. Second to be able to detect very early in a day perturbations in the network.

The first goal aims at obtaining representation of station that can then be used in other deep learning model to achieve other goals. This is why the unbiased nature of the representation is of the upmost importance.

By nature, the timeseries of frequentation of the subway networks are extremely regular with a week day/week-end period. However, some exterior factor (such as weather or public events) can impact this regularity. The goal of the project was to use the first hour of frequentation in the subway to predict how the network would be impacted for the day, allowing agents to adapt early to the situation.

To achieve these goals we first used a disentanglement deep learning model for generation of time series. By verifying with historical data that the generated time series were correct, we were able to confirm the solidity of our representations. We then used some clustering algorithms to enhances these representation and remove some bias. These representation were then used in another model designed to predict the frequentation of an entire day for the first few 15 minutes. The second model, which uses recent advances in forecasting, disentanglement and reccurent networks, was able to detect as early as 45 minutes into the day wether the day would be regular or not.