Machine learning with subsequent physics-based analytics for guiding transport planning

A methodology involving machine learning with subsequent physics-based method and a visualization map has been developed for enhancing environment-friendly transport planning. This methodology is applied for enabling predictions of traffic counts of bikes in Oslo city, with the physics-based models aiding in: (a) offering causality and interpretiveness for the ML predicted correlations (or patterns), and (b) more weather data for the trained machine learning to predict the traffic counts. In this work, we compare conventional ML methods (Artificial Neural network, Support Vector Machine (SVM), Random forest (RF)) for regression along with an unconventional use of sequential Long Short term memory (LSTM-RNN) based regression-formulation for predicting traffic counts of cycles at various traffic routes for both weather related and non-weather related independent variables. Amongst the various models compared, RF and LSTM model performs the best owing to nature of data. The RF quantifies the pattern that weather related variables are influencing cycle traffic count strongly, but ML being non-interpretive, it is not able to provide causation behind these correlations. A subsequent fluid dynamics-based analysis shows that weather influences thermal comfort level of cyclist and explains the observed ML predicted travel patterns. Finally, they are used together to provide a visualization map showing bike traffic count with the expectation to enable decision on infrastructure planning for the Oslo kommune (like a network of automated kiosk location). Such methodologies are a step towards encouraging better utilization of modern sharing-concepts (like, mobility-as-a-service and bike sharing).