Mengmeng HAN: Infusing deep-learning model with domain knowledge for better surface wind nowcasting
Infusing deep-learning model with domain knowledge for better surface wind nowcasting
[Context]
High temporal frequency wind speed and direction predictions are critical for short-term operational decisions in the aviation sector, including runway selection, aircraft arrival sequencing, and departure planning. These decisions benefit substantially from having reliable forecasts several hours in advance, allowing operators to evaluate potential scenarios and formulate robust strategies. This results in a demand for forecasts with long forecast horizons at the minute-level time scale, a challenging problem since wind fluctuations are highly random. In addition to accurate wind speed and direction at each nowcast step, accurate prediction of the occurrence and timing of wind change event is also significant, since wind change may lead to runway change or shutdown. However, predicting the step-wise wind direction and segment-wise wind change events are different and sometimes conflicting objectives.
[Methodology]
The present study aims at designing a deep-learning model that can predict wind direction and wind events at the same time with reasonable accuracy. Using temporal fusion transformer (TFT) as a backbone model, a side branch is specifically designed to capture signals relevant to wind change that are often sparse and coarser in temporal scale. Helped with domain knowledge from aviation forecasters, neighbouring stations close to the nowcast site and long NWP guidance are added as additional information for the model to extract information on wind regimes.
[Results]
Results show that an average improvement of 10%-30% can be achieved on wind speed and direction nowcast, 5% on wind event nowcasting on test datasets compared with baseline model. And case studies simulating real-time operational scenario also show earlier detection and better accuracy in timing.
[Audience]
The presentation will be of interest to ML model developers and audience with an interest in ML model architecture design and improvement.
[Keywords]
wind nowcast; probabilistic forecasting; wind speed; deep learning; weather forecasting