Type: Master's Assignment
Programme: Computer Science, Applied Mathematics, Embedded Systems, Sustainable Energy Technology
Contact: Gerwin Hoogsteen
Abstract
The future smart energy system needs to perform control actions continuously. Hereby coordination among devices happens both in space (i.e. which device does what) and time (e.g. when will the device execute this). Crucial to this are models to perform model predictive control and optimization. Currently, our optimization approaches already have a model.. These models are simplistic to make them mathematically usable for very efficient algorithms. However, such a simplistic model introduces errors in a real deployment with real devices. Creating a realistic model of each device is also not a realistic option, however.
With the advances in computing power, such a model based optimization approach may also be developed using machine learning. A model can be created from observations over time, e.g. it becomes a data driven model instead. The optimization itself is very much like machines that learn games, such as AlphaGo. With device optimization, choices (turns) must also be taken now and estimated for future intervals. Furthermore, a planning also results in a certain score (e.g. costs or deviations from the desired objective). Hence, we expect that machine learned device optimizations are possible.
Your task is to develop a machine learning based device optimization approach for a device. This can be a complete standalone algorithm, or may make use of one of our existing optimization approaches as a basis. Furthermore, exciting machine learning platforms can be integrated, for which a literature review needs to be performed. Performance and accuracy analysis must be carried out to compare the performance with existing solutions.