IAS Seminar "Engineering models and data science for wind energy"

Speaker:

Matthew Lennie, PhD, Institute of Fluid Dynamics and Technical Acoustics, TU Berlin

Abstract:

With the explosion of machine learning into the public consciousness, it may appear that many of the worlds challenging scientific problems will fall under the weight of vast computational power. While achievements such as AlphaStar are incredibly exciting and a worthy research effort, the usage of machine learning in other scientific fields can yield fantastic results with modest models and relatively small amounts of data. The truth is however, outside of the machine learning field, the possibilities of machine learning are poorly communicated. For example, the wind turbine industry is relatively unsophisticated with regards to machine learning, companies proudly announce that they have vast amounts of data, then quietly admit that they can’t access it with any kind of efficiency. They don’t attempt to use machine learning because they imagine that they need Silicon Valley scale datasets and computational resources. However, the opportunity here is that in many fields, we can make some fantastic wins with relatively modest approaches because relatively little has been attempted. In this talk, I will look at the traditional ways that these engineering tasks are tackled and show some highly efficient approaches for tackling problems with machine learning, i.e. identify aerodynamic features using a retrained RESNET model. This represents my approach to the applied machine learning side of the position at Jülich. On the more advanced side, I will give a brief overview of a current project where my team and I are building up a wind turbine controller using Reinforcement Learning.

Matthew Lennie was invited by Dr. Jenia Jitsev (JSC).