Demand model generation for shared mobility: a KDE approach

In recent times important changes on how people move around the cities have been happening. The expression “smart mobility” can well describe it. This definition includes: technology, infrastructures (park slots, charging stations, traffic signs, vehicles), mobility solutions and people. The aim is to lower the traffic and pollution, to create smart traffic flows with no interruption, and also boost the economies of scale in order to give everyone greater access to mobility. As a matter of fact, energy consumption is by far the greatest source of CO2 emissions, with 76% worldwide. This includes, among others, road transportation with 12.5%. In this optic, European Union has set the objective of reducing by 30% the emissions compared to 1990, by 2030. Furthermore, it has been proposed starting from 2035, an effective ban for new fossil-fuel cars, a clear signal to the car makers in order to accelerate their innovation on electric vehicles. In this context several business models inspired by sharing economy and Information Communication Technology grew up. Vehicle sharing is the rental of vehicles by the hour or by the minute as opposed to traditional day or week-long rentals. Members of the system have access to a fleet of vehicles that they can rent on an as-needed basis. The fee charged is based on the length of the rental in hours or minutes. With the so called free floating vehicle sharing, vehicles do not have home parking spaces but are instead can be parked anywhere within a city’s operating area. The ever-increasing installation of IoT object leads to a consequent generation of BigData. IoT and Big Data are strictly interconnected, originating a continuous cycle: data creation from IoT, data collection and analysis, with big data analytics pipelines, new configuration of the manufacturing and maintenance processes with the information extracted from data. In this thesis, we created a demand model that is capable to describe properly the free floating services users habits. The aim is to use this to extend an existing data-driven simulator named ODySSEUS, developed by Smart Data research group from which this work is supervised. The model is based on Poisson Processes for time domain, and a four dimension Kernel Density Estimate (KDE) for space. We use different datasets from the most important car sharing company in Europe, cleaning and applying different pre-processing steps on them, in order to be able to extract useful information. We concentrate in space domain and in particular in the optimization of the bandwidth parameter of KDE through machine learning cross validation approaches. We first investigate on the difference between discretized input and output and continuous one. We show the importance of optimizing properly the Fixed Bandwidth KDE, moreover we introduced Variable Bandwidth KDE approach. Results shows the difference between discretized input and output and continuous one, with different advantages with the continuous approach. Also, in terms of log-likelihood metric, the importance of optimizing the hyper-parameters bandwidth with machine learning cross validation approach. Yet, we illustrate how variable KDE can lead to some advantages in some cases.