Increasing accuracy and resolution of weather forecasts using deep generative models
Embodiments of the present invention provide the use of a conditional Generative Adversarial Network (GAN) to simultaneously correct and downscale (super-resolve) global ensemble weather or climate forecasts. Specifically, a generator deep neural network (G-DNN) in the cGAN comprises a corrector DNN (C-DNN) followed by a super-resolver DNN (SR-DNN). The C-DNN bias-corrects coarse, global meteorological forecasts, taking into account other relevant contextual meteorological fields. The SR-DNN downscales bias-corrected C-DNN output into G-DNN output at a higher target spatial resolution. The GAN is trained in three stages: C-DNN training, SR-DNN training, and overall GAN training, each using separate loss functions. Embodiments of the present invention significantly outperform an interpolation baseline, and approach the performance of operational regional high-resolution forecast models across an array of established probabilistic metrics. Crucially, embodiments of the present invention, once trained, produce high-resolution predictions in seconds on a single machine.
1. A method for training a generative adversarial network (GAN) for correcting and downscaling global numerical weather forecast data to generate an ensemble of forecast data for a target meteorological variable at a target spatial resolution, comprising:
pre-processing the global numerical weather forecast data to generate an ensemble of corrector input data at a first spatial resolution,
wherein the first spatial resolution is lower than the target spatial resolution,
wherein the GAN comprises a generator deep neural network (G-DNN) and a discriminator deep neural network (D-DNN),
wherein the G-DNN comprises a corrector deep neural network (C-DNN) followed by a super-resolver deep neural network (SR-DNN) having an output spatial resolution at the target spatial resolution, and
wherein the generated ensemble of corrector input data comprises data for the target meteorological variable and at least one context meteorological variable different from the target meteorological variable;
down-sampling observational data for the target meteorological variable from a second spatial resolution to the first spatial resolution, wherein the second spatial resolution is higher than the first spatial resolution, and wherein the observational data for the target meteorological variable corresponds to the ensemble of corrector input data over geopatch-time indices;
pre-training the C-DNN, using a first C-DNN loss function independent of the SR-DNN and D-DNN, wherein the first C-DNN loss function is computed based on a first C-DNN output generated from the ensemble of corrector input data, and the down-sampled observational data;
pre-training the SR-DNN, using a second SR-DNN loss function separate from the first C-DNN loss function and independent of the D-DNN, wherein the second SR-DNN loss function is computed based on the first C-DNN output, the down-sampled observational data, a first SR-DNN output generated from the first C-DNN output, and the observational data for the target meteorological variable; and
training the GAN, using a third D-DNN loss function computed based on a second C-DNN output generated from the ensemble of corrector input data and a random vector input.
2. The method of claim 1 , further comprising:
generating the ensemble of forecast data for the target meteorological variable at the target spatial resolution, using the G-DNN.
3. The method of claim 1 , further comprising:
generating a probabilistic forecast from the ensemble of forecast data for the target meteorological variable at the target spatial resolution.
4. The method of claim 1 , wherein the target meteorological variable is precipitation accumulation, and wherein the at least one context meteorological variable is surface temperature, convective available potential energy, convective inhibition, or column water.
5. The method of claim 1 , wherein the first loss C-DNN function is computed further based on a fractions skill score generated from the first C-DNN output and the down-sampled observational data.
6. The method of claim 1 , wherein the third D-DNN loss function is computed further based on the down-sampled observational data.
7. The method of claim 1 , further comprising:
weighted sampling the ensemble of corrector input data and the observational data.
8. The method of claim 1 , wherein the C-DNN has a same input and output spatial resolution, and wherein the SR-DNN has an input spatial resolution at the first spatial resolution.
9. A system for training a generative adversarial network (GAN) for correcting and downscaling global numerical weather forecast data to generate an ensemble of forecast data for a target meteorological variable at a target spatial resolution, comprising:
at least one processor; and
a non-transitory physical storage medium for storing program code and accessible by the processor, the program code when executed by the processor causes the processor to:
pre-process the global numerical weather forecast data to generate an ensemble of corrector input data at a first spatial resolution,
wherein the first spatial resolution is lower than the target spatial resolution,
wherein the GAN comprises a generator deep neural network (G-DNN) and a discriminator deep neural network (D-DNN),
wherein the G-DNN comprises a corrector deep neural network (C-DNN) followed by a super-resolver deep neural network (SR-DNN) having an output spatial resolution at the target spatial resolution, and
wherein the generated ensemble of corrector input data comprises data for the target meteorological variable and at least one context meteorological variable different from the target meteorological variable;
down-sample observational data for the target meteorological variable from a second spatial resolution to the first spatial resolution, wherein the second spatial resolution is higher than the first spatial resolution, and wherein the observational data for the target meteorological variable corresponds to the ensemble of corrector input data over geopatch-time indices;
pre-train the C-DNN, using a first C-DNN loss function independent of the SR-DNN and D-DNN, wherein the first C-DNN loss function is computed based on a first C-DNN output generated from the ensemble of corrector input data, and the down-sampled observational data;
pre-train the SR-DNN, using a second SR-DNN loss function separate from the first C-DNN loss function and independent of the D-DNN, wherein the second SR-DNN loss function is computed based on the first C-DNN output, the down-sampled observational data, a first SR-DNN output generated from the first C-DNN output, and the observational data for the target meteorological variable; and
train the GAN, using a third D-DNN loss function computed based on a second C-DNN output generated from the ensemble of corrector input data and a random vector input.
10. The system of claim 9 , wherein the program code when executed by the processor further causes the processor to:
generate the ensemble of forecast data for the target meteorological variable at the target spatial resolution, using the G-DNN.
11. The system of claim 9 , wherein the program code when executed by the processor further causes the processor to:
generate a probabilistic forecast from the ensemble of forecast data for the target meteorological variable at the target spatial resolution.
12. The system of claim 9 , wherein the target meteorological variable is precipitation accumulation, and wherein the at least one context meteorological variable is surface temperature, convective available potential energy, convective inhibition, or column water.
13. The system of claim 9 , wherein the first C-DNN loss function is computed further based on a fractions skill score generated from the first C-DNN output and the down-sampled observational data.
14. The system of claim 9 , wherein the third D-DNN loss function is computed further based on the down-sampled observational data.
15. The system of claim 9 , wherein the program code when executed by the processor further causes the processor to:
weighted sample the ensemble of corrector input data and the observational data.
16. The system of claim 9 , wherein the C-DNN has a same input and output spatial resolution, and wherein the SR-DNN has an input spatial resolution at the first spatial resolution.
17. A non-transitory physical storage medium for training a generative adversarial network (GAN) for correcting and downscaling global numerical weather forecast data to generate an ensemble of forecast data for a target meteorological variable at a target spatial resolution, the storage medium comprising program code stored thereon, that when executed by a processor causes the processor to:
pre-process the global numerical weather forecast data to generate an ensemble of corrector input data at a first spatial resolution,
wherein the first spatial resolution is lower than the target spatial resolution,
wherein the GAN comprises a generator deep neural network (G-DNN) and a discriminator deep neural network (D-DNN),
wherein the G-DNN comprises a corrector deep neural network (C-DNN) followed by a super-resolver deep neural network (SR-DNN) having an output spatial resolution at the target spatial resolution, and
wherein the generated ensemble of corrector input data comprises data for the target meteorological variable and at least one context meteorological variable different from the target meteorological variable;
down-sample observational data for the target meteorological variable from a second spatial resolution to the first spatial resolution, wherein the second spatial resolution is higher than the first spatial resolution, and wherein the observational data for the target meteorological variable corresponds to the ensemble of corrector input data over geopatch-time indices;
pre-train the C-DNN, using a first C-DNN loss function independent of the SR-DNN and D-DNN, wherein the first C-DNN loss function is computed based on a first C-DNN output generated from the ensemble of corrector input data, and the down-sampled observational data;
pre-train the SR-DNN, using a second SR-DNN loss function separate from the first C-DNN loss function and independent of the D-DNN, wherein the second SR-DNN loss function is computed based on the first C-DNN output, the down-sampled observational data, a first SR-DNN output generated from the first C-DNN output, and the observational data for the target meteorological variable; and
train the GAN, using a third D-DNN loss function computed based on a second C-DNN output generated from the ensemble of corrector input data and a random vector input.
18. The non-transitory physical storage medium of claim 17 , wherein the target meteorological variable is precipitation accumulation, and wherein the at least one context meteorological variable is surface temperature, convective available potential energy, convective inhibition, or column water.
19. The non-transitory physical storage medium of claim 17 , wherein the first C-DNN loss function is computed further based on a fractions skill score generated from the first C-DNN output and the down-sampled observational data.
20. The non-transitory physical storage medium of claim 17 , wherein the third D-DNN loss function is computed further based on the down-sampled observational data.