Spatio-temporal algorithms for predicting the usage of bike-sharing systems

With the recent concerns about climate change and the tendency to promote smart mobility systems, there is an increasing interest on shared means of transport around the world, such as Bike Sharing Systems (BSS), which constitute a valid green alternative for movements within cities. Given these circumstances, many studies have been carried out by the scientific community to improve the service of Bike Sharing Systems focusing on customer satisfaction, by finding optimal locations for bike stations, studying mobility patterns in the cities of interest, enhancing bike redistribution among stations, offering customers predictions in the next minutes about bike availability in the stations. This thesis aims at contributing at the last two questions by proposing predictive models to forecast in the short term the number of bikes available and free slots in each station of the Barcelona BSS, for example in 20 minutes or an hour. The proposed solution exploits stations’ behavior in the previous time instants expressed in terms of the number of available bikes, the number of free slots and the dimension of each station, whether the station is full or empty, as well as the spatial relationships between the stations by means of some spatial-temporal features. The situation of nearby stations was taken into account by computing how many neighbors stations were full or empty and which was the average station occupation in the neighborhood in the previous timestamps. Also a distinction between working days or holidays was made which has been helpful in the analysis. A windowing technique, the sliding window method, made it possible to consider the observations in the previous instants of time for all the features, transforming multivariate time series into tables, thus making the forecasting problem a supervised learning problem such that it was possible to use regression algorithms in the task, like linear regression and Support Vector Regression. Three versions of this predictive system were also tested: 1) it was decided to combine each regression algorithm with the baseline, by using regression in non-stationary moments and the baseline model, that predicts the current value in the future, in stationary moments. In this setting, the variance threshold that allows to distinguish stationary windows from non-stationary ones was a parameter to consider; 2) another test was done by predicting the number of free slots in each station, rather than the number of bikes available; 3) it was considered the addition of a new feature to identify weekly patterns, which consists of the difference in the number of bikes in the previous week in the same time slot. To evaluate model performances, R2 and RMSE metrics were compared with those of the baseline, exceeding it. Further comparisons were also made between the different versions of the prediction system with the aim of understanding the effects of changing the algorithm used, the variance threshold, the size of the sliding window and the forecast horizon.