Semi-supervised deep model for turbulence forecasting
A method for employing a unified semi-supervised deep learning (DL) framework for turbulence forecasting is presented. The method includes extracting historical and forecasted weather features of a spatial region, calculating turbulence indexes to fill feature cubes, each feature cube representing a grid-based 3D region, and building an encoder-decoder framework based on convolutional long short-term memory (ConvLSTM) to model spatio-temporal correlations or patterns causing turbulence. The method further includes employing a dual label guessing component to dynamically integrate complementary signals from a turbulence forecasting network and a turbulence detection network to generate pseudo-labels, reweighing the generated pseudo-labels by a heuristic label quality detector based on KL-Divergence, applying a hybrid loss function to predict turbulence conditions, and generating a turbulence dataset including the predicted turbulence conditions.
1. A computer-implemented method executed on a processor for employing a unified semi-supervised deep learning (DL) framework for turbulence forecasting, the method comprising:
extracting historical and forecasted weather features of a spatial region;
calculating turbulence indexes from the extracted historical and forecasted weather features to fill feature cubes, each feature cube representing a grid-based 3D region;
building an encoder-decoder framework taking as input the feature cubes, and based on convolutional long short-term memory (ConvLSTM) to model spatio-temporal correlations or patterns causing turbulence;
employing a dual label guessing component to dynamically integrate complementary signals from a turbulence forecasting network and a turbulence detection network to generate pseudo-labels;
reweighing the generated pseudo-labels by a heuristic label quality detector based on KL-Divergence;
applying a hybrid loss function to the pseudo-labels generated by the dual label guessing component to predict turbulence conditions; and
generating a turbulence dataset including the predicted turbulence conditions.
2. The method of claim 1 , wherein forecasted turbulence feature cubes are fed into the turbulence detection network.
3. The method of claim 2 , wherein the turbulence detection network outputs detected turbulence level cubes.
4. The method of claim 1 , wherein the pseudo-labels are generated for unlabeled grids based on the turbulence forecasting network and the turbulence detection network, are coupled with corresponding features as ground truth, and added to training data included in the turbulence dataset.
5. The method of claim 1 , wherein the heuristic label quality detector calculates a quality score for each grid.
6. The method of claim 1 , wherein the hybrid loss function includes a supervised loss for labeled grids and an unsupervised loss for grids with pseudo-labels.
7. The method of claim 1 , wherein the turbulence forecasting network and the turbulence detection network are co-trained in a semi-supervised manner.
8. A non-transitory computer-readable storage medium comprising a computer-readable program for employing a unified semi-supervised deep learning (DL) framework for turbulence forecasting, wherein the computer-readable program when executed on a computer causes the computer to perform the steps of:
extracting historical and forecasted weather features of a spatial region;
calculating turbulence indexes from the extracted historical and forecasted weather features to fill feature cubes, each feature cube representing a grid-based 3D region;
building an encoder-decoder framework taking as input the feature cubes, and based on convolutional long short-term memory (ConvLSTM) to model spatio-temporal correlations or patterns causing turbulence;
employing a dual label guessing component to dynamically integrate complementary signals from a turbulence forecasting network and a turbulence detection network to generate pseudo-labels;
reweighing the generated pseudo-labels by a heuristic label quality detector based on KL-Divergence;
applying a hybrid loss function to the pseudo-labels generated by the dual label guessing component to predict turbulence conditions; and
generating a turbulence dataset including the predicted turbulence conditions.
9. The non-transitory computer-readable storage medium of claim 8 , wherein forecasted turbulence feature cubes are fed into the turbulence detection network.
10. The non-transitory computer-readable storage medium of claim 9 , wherein the turbulence detection network outputs detected turbulence level cubes.
11. The non-transitory computer-readable storage medium of claim 8 , wherein the pseudo-labels are generated for unlabeled grids based on the turbulence forecasting network and the turbulence detection network, are coupled with corresponding features as ground truth, and added to training data included in the turbulence dataset.
12. The non-transitory computer-readable storage medium of claim 8 , wherein the heuristic label quality detector calculates a quality score for each grid.
13. The non-transitory computer-readable storage medium of claim 8 , wherein the hybrid loss function includes a supervised loss for labeled grids and an unsupervised loss for grids with pseudo-labels.
14. The non-transitory computer-readable storage medium of claim 8 , wherein the turbulence forecasting network and the turbulence detection network are co-trained in a semi-supervised manner.
15. A system for employing a unified semi-supervised deep learning (DL) framework for turbulence forecasting, the system comprising:
a memory; and
one or more processors in communication with the memory configured to:
extract historical and forecasted weather features of a spatial region;
calculate turbulence indexes from the extracted historical and forecasted weather features to fill feature cubes, each feature cube representing a grid-based 3D region;
build an encoder-decoder framework taking as input the feature cubes, and based on convolutional long short-term memory (ConvLSTM) to model spatio-temporal correlations or patterns causing turbulence;
employ a dual label guessing component to dynamically integrate complementary signals from a turbulence forecasting network and a turbulence detection network to generate pseudo-labels;
reweigh the generated pseudo-labels by a heuristic label quality detector based on KL-Divergence;
apply a hybrid loss function to the pseudo-labels generated by the dual label guessing component to predict turbulence conditions; and
generate a turbulence dataset including the predicted turbulence conditions.
16. The system of claim 15 , wherein forecasted turbulence feature cubes are fed into the turbulence detection network.
17. The system of claim 16 , wherein the turbulence detection network outputs detected turbulence level cubes.
18. The system of claim 15 , wherein the pseudo-labels are generated for unlabeled grids based on the turbulence forecasting network and the turbulence detection network, are coupled with corresponding features as ground truth, and added to training data included in the turbulence dataset.
19. The system of claim 15 , wherein the heuristic label quality detector calculates a quality score for each grid.
20. The system of claim 15 , wherein the hybrid loss function includes a supervised loss for labeled grids and an unsupervised loss for grids with pseudo-labels.