(Georgia Institute of Technology, 2021-01)
Kim, Junghyun; Zhang, Chao; Mavris, Dimitri N.; Briceno, Simon
Many research groups have been committed to developing numerical models for weather forecasts. The models are currently used to predict weather patterns and trends in the aviation industry. In particular, pilots receive wind information predicted by the models and use the forecast to not only calculate how much fuel is needed for a flight but also optimize flight routes by seeking favorable winds. One potential issue is that the models provide relatively coarse wind information in both space and time, which potentially leads to inaccurate calculation of fuel consumption. This research aims to yield a continuous wind prediction model by combining a supervised learning algorithm with the Inverse Distance Weighting technique. Specifically, this research compares three different supervised learning algorithms that include Gaussian Process, Multi-Layer Perceptron, and Support Vector Machine to identify the most appropriate algorithm. The selected algorithm is then compared to a linear interpolation method that is widely used in current flight planning systems for obtaining continuous wind information. A case study is performed with the real Delta Airlines flight 1944 to evaluate the proposed methodology. The results show that 1) the Support Vector Machine provides a better wind prediction compared to the other models, 2) the supervised learning-based regression method performs better than the linear interpolation method in wind predictions, and 3) there are 16 seconds of difference between the real flight (12,117 seconds) and the simulated flight (12,101 seconds) for the cruise portion, indicating that the proposed methodology generates valid results as long as input wind data is provided accurately.