BusGrid is implemented as a series of interconnected modules. The data collection module gathers a series of data from the AVL and APC systems and other installed sensors such as the engine ignition on/off, door sensors etc. Using the bus’s wireless communication system (e.g. GSM), a unified data-frame containing the sensor data is sent to the database server in real-time. Next, a data layer is created by preprocessing the raw data of the database. This stage includes data cleaning, regularisation and aggregation of data (e.g. creating a per-day view of the data). The data-layer created, forms the data-sets used by the modules that calculate the KPIs of the bus routes and by the Machine Learning modules. The KPIs are calculated directly from the raw data.

The Machine Learning modules are used for the prediction of critical KPIs such as passenger demand and for the adaptation of the system in various cases. The first Machine Learning module (Prediction of Demand) estimates the demand for each specific bus route at each bus stop. It uses examples of temporal and geospatial data, combined with the historical data from the APC sensors, for the supervised learning of a regression model which predicts people demand for a specific bus stop, bus and period of time. The average waiting time module produces a KPI based on the actual average waiting time weighted by the passenger demand (number of people waiting) so that waiting time is expressed in man-hours.

Finally, the RL-BUS module uses Reinforcement Learning methods to dynamically create adaptive bus schedules optimised for bus productivity and quality of service. It is a semi-supervised model, learning a scheduling policy for dispatching buses on the correct time to achieve optimal bus productivity with respect to the average bus fullness and waiting time (quality of service indicators). The KPIs calculation modules and the Machine Learning modules are developed as software libraries, separating the interface tier of the system (GUI) from its business logic.