FutureFlow: A Deep Learning Approach to Short-Term Wind Gust Prediction and Control, Dylan Epstein-Gross, UG '25 (BFB85A86)

From Research Princeton Research Day May 1st, 2024

70 plays You unliked the media.

My project explores the application of various deep learning methods to the problem of short-term wind gust prediction central to wind-aware control on real-world quadrotors, a ubiquitous issue for safety and energy efficiency. A variety of models, from simple to state-of-the-art – Linear Regressions, Deep Neural Networks, Recurrent Neural Networks, Long Short-Term Memory Networks, Time Series Transformers, and ANN-Markov Models – are trained and tested on both synthetic and real-world wind data, with the most successful employed in a basic simulator to evaluate their real-world feasibilities. Results show that the models significantly outperform baselines in predicting the complex behavior of wind, with the most effective ones also reducing positional error within the simulator. These conclusions warrant future hardware implementation of the models, potentially on
the Intelligent Robot Motion Lab’s FlowDrone, to create the novel FutureFlow.

Date: May 3rd, 2024
Speaker: Dylan Epstein-Gross, UG '25 (BFB85A86)
Department: Computer Science
Appears In: Princeton Research Day 2024 Research Day

Looking for more...

FutureFlow: A Deep Learning Approach to Short-Term Wind Gust Prediction and Control, Dylan Epstein-Gross, UG '25 (BFB85A86)

Related Media