Background and activities

Current research

My main current research concerns safe Reinforcement Learning and Data-Driven MPC. I currently have 4 PhD students and one postdoc working on this topic. This work is strongly connected to the industry, with large industrial partners such as DNV GL and Konsberg Marine. I have an additional PhD student working on stochastic MPC techniques for mission-wide safety guarantees.

I also co-supervise several PhD projects related to the Artificial Pancreas problem, multi-rotor wind turbine control. Finally, I supervise and cosupervise several industrial PhD projects based at Volvo (Sweden), focusing on powertrain control, 2nd life of batteries, and traffic management. 

MSc thesis projects for 2019-2020

Probabilistic Modelling of Micro-Grid Operation

Sintef Digital is currently involved in several industrial research collaborations within the energy sector. Of particular interest is migro-grid Energy Management Systems (EMS). A micro-grid is an interconnected grid of loads and energy storage, fed by distributed and renewable energy sources. The micro-grid may be operated in island-mode or connected to the grid. The goal of the EMS is to optimize the cost of energy while serving the loads and managing the availability of energy. Varying and uncertain loads, weather conditions, and price of electricity makes this optimization non-trivial. In this project we aim at gaining a better insight into the varying disturbances that affect the optimal EMS operation. Predictive probabilistic models of these disturbances are crucial for the EMS performance. Our goal is to develop in-house knowledge of such predictive models: how to build them efficiently, with the aim of being used in the context of optimal control and Model Predictive Control.

NTNU ITK supervisor: Sebastien Gros (sebastien.gros@ntnu.no), 

Industrial supervisor: Dr. Phillip Maree, SINTEF Digital (Phillip.maree@sintef.no)

 

Data-driven MPC of Micro-Grid Operation

Sintef Digital is currently involved in several industrial research collaborations within the energy sector. Of particular interest is migro-grid Energy Management Systems (EMS). A micro-grid is an interconnected grid of loads and energy storage, fed by distributed and renewable energy sources. The micro-grid may be operated in island-mode or connected to the grid. The goal of the EMS is to optimize the cost of energy while serving the loads and managing the availability of energy. Varying and uncertain loads, weather conditions, and price of electricity makes this optimization non-trivial. In this project we aim at experimenting with data-driven optimal control techniques, including Model Predictive Control (MPC) including advanced statistical modelling and Reinforcement Learning (RL), for the control and optimization of micro-grids. Our goal is to develop in-house knowledge of how to use data in the most efficient way for micro-grid control.

NTNU ITK supervisor: Sebastien Gros (sebastien.gros@ntnu.no), 

Industrial supervisor: Dr. Phillip Maree, SINTEF Digital (Phillip.maree@sintef.no)

 

Reinforcement Learning with Safety guarantees

Reinforcement Learning (RL) is a powerful tool to generate optimal control policies from data for systems that are difficult to model. However, imposing safety constraints and requirements on RL-based policies is still a difficult task. In this project, we aim at exploring novel ideas and techniques to tackle this issue. DNVGL is actively performing industrial research on the use of RL for marine applications. DNVGL is testing RL solutions on the Revolte platform. RL-based policies have been developed, but delivering safety certificates for these policies is very difficult. Indeed, as a data-driven technique, safety in pure RL-based control can only emerge from collecting a large amount of experiments where safe operations are breached. In this project, we are interested in improving on that approach, and in particular in using model and knowledge-based techniques to introduce safety in the baseline RL algorithms used at DNVGL. The project will focus on the REVOLTE simulated and experimental platform.  

NTNU ITK supervisor: Sebastien Gros (sebastien.gros@ntnu.no), 

Industrial supervisor: Jon Arne Glomsrund (Jon.Arne.Glomsrund@dnvgl.com)

 

Mission-wide Stochastic MPC

Stochastic and Robust Model Predictive Control (MPC) allow for taking model uncertainties and random disturbances from the environment into account when calculating control policies. Robust MPC considers worst-case scenarios, while Stochastic MPC aims at limiting the probability of violating the constraints at any time instant below a prescribed value. Both Robust and Stochastic MPC have been and still are intensively investigated in research. A deficiency of Robust MPC is that it - because it adopts a worst-case point of view - can be very conservative, and does not necessarily offer a control solution. Stochastic MPC solves these issues, but it neglects the correlation in time between violating the constraints, and as a result, it offers an incorrect representation of the risks incurred by a system over a mission. Novel research attempts the development of Stochastic MPC schemes where the overall probability of a system violating some critical constraints over a mission is considered, rather than at each time instant. The application and further development of this research is an on-going work. In particular, an extensive testing and improvement of the algorithmic principle allowing deploying these ideas in practice is required. 

NTNU ITK supervisor: Sebastien Gros (Sebastien.gros@ntnu.no)

Scientific, academic and artistic work

Journal publications