The ever increasing market penetration of electrified vehicles, the diffusion of intermittent renewable energy sources, together with the growing public interest in environmental issues, are acting as a catalyst for research on smart-grid energy management. It is nowadays widely acknowledged that the growing number of electrified vehicles can cause very high peak electricity demand, which needs to be properly addressed and managed in order to avoid electricity demand/supply unbalance and system breakdown.
Digital Science and Technology
Upcoming Energy transition should increase the level of investment in production/storage of renewable energy systems distributed within the electricity network. These investments are usually profitable after several years and their level of profitability is quite uncertain due to the stochasticity and the intermittency of renewable production.
Many fields are currently benefiting from the development of machine learning methods, and Computational Fluid Dynamics (CFD) is no exception. In this context, an important opportunity is the possibility of speeding up calculations by replacing expensive numerical solvers with equivalent statistical models, based on machine learning methods. The key benefit is the possibility of performing more precise calculations thanks to the inclusion of more detailed physico-chemical phenomena, which is made possible by the gain in computation time.
The stochastic gradient method (SGD) is the currently prevailing technology for training neural networks. This method takes advantage of the specific structure of the cost function to be minimized for which the gradient is sought. Compared to a classical descent method, the calculation of the true gradient, which is an average over the number of data, is replaced by a random element of the sum, hence the name stochastic gradient.
Pollutant emissions related to transport applications have various origins and types. Measurements by sampling permit very localized estimates of concentrations and to have maps at an urban scale by using dispersion models. On the one hand present simulators use simplified representations of atmospheric turbulence that can be discussed when complex situations are considered (irregular buildings heights or dispositions generating strong 3D turbulent effects).
Describing natural processes that control our environment is a major challenge arising in numerous domains. In particular, describing long term erodability is mandatory to understand the impact of climate change on coastal landscapes. Such areas being among the most populated in the world, the social and economic issues are tremendous. Waves dynamic, mostly controlled by their frequency and amplitude, is the major engine of coastal erosion. Slope instabilities and coastal landslides depend on their frequency and amplitude, as sediments transport from land to sea.
Eco-routing is a driving assistance strategy for connected and/or automated vehicles aimed at suggesting the most energy-efficient route from an origin to a destination at a given time. Such a control strategy is often devised for individual vehicles in road networks subject to time-varying traffic conditions.
The heat generated by the electric motor is due to losses within multiple components like stator windings and rotor magnets or conductors. Among the different cooling strategies, we focus this study on an active convective cooling technology, where heat is removed from the motor and transferred to another location to eventually be rejected to the ambient environment. Heat may be extracted through active cooling such as a cooling jacket or a spray injecting directly a cold liquid on the stator or the rotor.
Numerical simulation is an essential tool in order to model complex physical processes. Modern parallel codes implement complex algorithms that solve large systems of non-linear or partial differential equations. These computations produce a large amount of data that is usually discarded for future similar computations. The aim of this thesis is to leverage these data using machine learning in order to improve numerical simulation performance.