Numerics and Software


Fault tolerant control using Gaussian processes on FPGA

Gaussian Processes – a machine learning method, is gaining attention for identifying nonlinear dynamical systems due to numerous advantages. However, the Gaussian processes are a computationally expensive tool which limits their application for real-time control and/or embedded control systems. 


In this work, we will deploy Gaussian processes on FPGAs using the LAfF – a code generation toolbox developed at Automatic Control Laboratory. We will consider model predictive control coupled with Gaussian processes for fault tolerant control with case studies of various flight control scenarios.


Prerequisite: C programming skills are required. Familiarity with FPGAs, Gaussian Processes, and Model Predictive control is desired but not required. 


Keywords: Machine learning, Field Programmable Gate Arrays, Model Predictive Control, Gaussian Processes, Embedded systems

Professor: Colin Jones
Type of project: Semester or Master
Contact: Harsh Shukla


Development of a software library for microgrid simulation

In the past decades, there has been a paradigm shift in the operation of power systems with an increased focus on localized and distributed generation as opposed to centralised mechanisms. Microgrids (mGs), both AC and DC, are spatially distributed systems composed of multiple small subsystems, for example, fexible loads, distributed small-scale generation, storage units etc., interconnected to each other through an electrical network. The manifold advantages of mGs like enhanced power quality, reduced transmission losses, capability to operate in grid-connected and islanded modes, and compatibility with renewable distributed generation, make them a promising operational architecture for future power systems. Moreover, recent series of legislation, strongly favouring the concept of sustainable and renewable energy, has reinvigorated interest in stable operation and control of mGs.

In order to study mGs and successively apply efficient control algorithms, a modelling framework is of paramount importance. The goal of this project is to ease simulation and enable fast prototyping of AC and DC mGs. Some basic components are present in simulation environments for electrical systems such as PSCAD, which not only allow one to build custom libraries of blocks for facilitating the modeling of specific electric systems, but also present a graphical user interface. However, these are general-purpose software packages, and building realistic models of large mGs can be time consuming. In this project, we will develop libraries collecting models of DGUs that are commonly found in AC and DC mGs. PSCAD is chosen as the software environment for mG simulation. Theblocks for controllers that will be developed as part of this project will be coded from scratch, and will eventually be made available publicly.

Requirements: No specific prior requirements. Familiarity with PSCAD, and mGs is a plus.

Professor: Giancarlo Ferrari Trecate
Type of project: Semester or Master
Contact: Pulkit Nahata


Exploration of highly parallelizable optimisation algorithms for controlThe figure is borrowed from Dattorro: Convex Optimization and Euclidean Distance Geometry

The need for real-time high-speed model predictive control (MPC) is evident during the last years, dictated from advanced control applications and industrial processes. Even more recently, the trend for distributed control policies has raised attention, motivated either from problems that maintain a distributed structural representation or from the need for decentralized computations.  

Decomposition algorithms seem to be the solution to many of the above problems. There is a vast and rapidly growing literature in these methods that have recently made their way into the control community.

The aim of the project is to explore a few very recent algorithms that seem to achieve a high level of parallelisation given a specific problem. The student is expected to understand the theoretical content of the methods, to apply them to MPC problems and to perform comparisons. Ideally, an efficient C-code implementation would be the project’s output.


Remark: The figure is borrowed from Dattorro: Convex Optimization and Euclidean Distance Geometry


Professor: Colin Jones
Type of project: Master or Semester
Contact: Georgios Stathopoulos


Collaborative tracking using real-time distributed optimization

The Fast Toy Lab is interested in fast optimisation-based control techniques for vehicles. We have recently been developing an algorithm for distributed real-time nonlinear optimal control.
As simulation results are promising, we are intending to test our algorithm on hardware. The goal of this project is to improve and implement our strategy on a collaborative tracking problem, where a leading vehicle follows a path while followers keep a specific formation. This can be formulated as a nonlinear optimal control problem to be solved online. The first steps of the project will be based on simulation, then the proposed algorithm should be validated and implemented in C.

Requirements: Model Predictive Control/Optimization, C, Matlab

Professor: Colin Jones
Type of project: Master 
Contact: Giorgos Stathopoulos