Systems and methods of data preprocessing and augmentation for neural network climate forecasting models
Methods and systems for training a neural network (NN)-based climate forecasting model on a pre-processed multi-model ensemble of global climate simulation data from a plurality of global climate simulation models (GCMs), are disclosed. The methods and systems perform steps of determining a common spatial scale and a common temporal scale for the multi-model ensemble of global climate simulation data; spatially re-gridding the multi-model ensemble to the common spatial scale; temporally homogenizing the multi-model ensemble to the common temporal scale; augmenting the spatially re-gridded, temporally homogenized multi-model ensemble with synthetic simulation data generated from the spatially re-gridded, temporally homogenized multi-model ensemble; and training the NN-based climate forecasting model using the spatially re-gridded, temporally homogenized, and augmented multi-model ensemble of global climate simulation data. Embodiments of the present invention enable accurate climate forecasting without the need to run new dynamical global climate simulations on supercomputers.
1. A method for training a neural network (NN)-based climate forecasting model, comprising:
generating a multi-model ensemble of global climate simulation data by combining simulation data generated from at least two of a plurality of global climate models (GCMs);
determining a common spatial scale and a common temporal scale for the multi-model ensemble of global climate simulation data;
spatially re-gridding the multi-model ensemble of global climate simulation data to the common spatial scale;
temporally homogenizing the multi-model ensemble of global climate simulation data to the common temporal scale;
augmenting the spatially re-gridded, temporally homogenized multi-model ensemble of global climate simulation data with synthetic simulation data, to generate a spatially re-gridded, temporally homogenized, and augmented multi-model ensemble of global climate simulation data, wherein the synthetic simulation data is generated from the spatially re-gridded, temporally homogenized multi-model ensemble of global climate simulation data; and
training the NN-based climate forecasting model using the spatially re-gridded, temporally homogenized, and augmented multi-model ensemble of global climate simulation data, wherein the NN-based climate forecasting model comprises a predictive neural network, wherein for a given input into the NN-based climate forecasting model, a forecast output at a target lead time is generated by the NN-based climate forecasting model based on the given input, wherein the forecast output at the target lead time is compared to a corresponding desired output at the target lead time during the training, and wherein the given input and the corresponding desired output at the target lead time are selected from the multi-model ensemble of global climate simulation data.
2. The method of claim 1 , wherein the spatially re-gridding the multi-model ensemble of global climate simulation data comprises at least one of spatial interpolation and spatial extrapolation to achieve a common spatial resolution.
3. The method of claim 1 , further comprising:
computing a forecast skill score for the multi-model ensemble of global climate simulation data, based on a data predictor function, the NN-based climate forecasting model, or a model analog of the NN-based climate forecasting model; and
updating the common spatial scale based on the forecast skill score.
4. The method of claim 1 , wherein the common temporal scale comprises a common date system, and wherein the temporally homogenizing the multi-model ensemble of global climate simulation data comprises converting the multi-model ensemble of global climate simulation data to the common date system.
5. The method of claim 1 , wherein the temporally homogenizing the multi-model ensemble of global climate simulation data comprises at least one of temporal interpolation and temporal extrapolation to achieve a common temporal resolution.
6. The method of claim 1 , wherein the spatially re-gridded, temporally homogenized multi-model ensemble of global climate simulation data comprises temperature data, and wherein the synthetic simulation data is generated from the spatially re-gridded, temporally homogenized multi-model ensemble of global climate simulation data through a statistical augmentation process comprising:
decomposing the temperature data into a first plurality of signal components, wherein a first signal component is related to El Nino-Southern Oscillation (ENSO), a second signal component is related to Atlantic Multidecadal Oscillation (AMO), a third signal component is related to Pacific Decadal Oscillation (PDO), and a fourth signal component is related to global warming;
generating a second plurality of surrogate time series from the first plurality of signal components, wherein at least one surrogate time series is generated by randomizing a phase of one of the first, second, and third signal components in a frequency domain;
generating a surrogate internal variability time series by resampling historical observational internal variability data; and
generating the synthetic simulation data based on the second plurality of surrogate time series, and the surrogate internal variability time series.
7. The method of claim 1 , wherein augmenting the spatially re-gridded, temporally homogenized multi-model ensemble of global climate simulation data comprises land-sea masking.
8. The method of claim 1 , wherein the predictive neural network is selected from the group consisting of a Convolutional Recurrent Neural Network (CRNN) having at least one Long Short-Term Memory (LSTM) layer, a Spherical Convolutional Neural Network (S 2 -CNN), and a transformer neural network.
9. The method of claim 1 , further comprising:
forecasting at least one target climate variable at the target lead time using the NN-based climate forecasting model.
10. A system for training a neural network (NN)-based climate forecasting model, 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:
generate a multi-model ensemble of global climate simulation data by combining simulation data generated from at least two of a plurality of global climate models (GCMs);
determine a common spatial scale and a common temporal scale for the multi-model ensemble of global climate simulation data;
spatially re-grid the multi-model ensemble of global climate simulation data to the common spatial scale;
temporally homogenize the multi-model ensemble of global climate simulation data to the common temporal scale;
augment the spatially re-gridded, temporally homogenized multi-model ensemble of global climate simulation data with synthetic simulation data, to generate a spatially re-gridded, temporally homogenized, and augmented multi-model ensemble of global climate simulation data, wherein the synthetic simulation data is generated from the spatially re-gridded, temporally homogenized multi-model ensemble of global climate simulation data; and
train the NN-based climate forecasting model using the spatially re-gridded, temporally homogenized, and augmented multi-model ensemble of global climate simulation data, wherein the NN-based climate forecasting model comprises a predictive neural network, wherein for a given input into the NN-based climate forecasting model, a forecast output at a target lead time is generated by the NN-based climate forecasting model based on the given input, wherein the forecast output at the target lead time is compared to a corresponding desired output at the target lead time during the training, and wherein the given input and the corresponding desired output at the target lead time are selected from the multi-model ensemble of global climate simulation data.
11. The system of claim 10 , wherein the spatially re-grid the multi-model ensemble of global climate simulation data comprises at least one of spatial interpolation and spatial extrapolation to achieve a common spatial resolution.
12. The system of claim 10 , wherein the program code when executed by the processor further causes the processor to:
compute a forecast skill score for the multi-model ensemble of global climate simulation data, based on a data predictor function, the NN-based climate forecasting model, or a model analog of the NN-based climate forecasting model; and
update the common spatial scale based on the forecast skill score.
13. The system of claim 10 , wherein the common temporal scale comprises a common date system, and wherein the temporally homogenize the multi-model ensemble of global climate simulation data comprises converting the multi-model ensemble of global climate simulation data to the common date system.
14. The system of claim 10 , wherein the temporally homogenize the multi-model ensemble of global climate simulation data comprises at least one of temporal interpolation and temporal extrapolation to achieve a common temporal resolution.
15. The system of claim 10 , wherein the spatially re-gridded, temporally homogenized multi-model ensemble of global climate simulation data comprises temperature data, and wherein the program code when executed by the processor further causes the processor to generate the synthetic simulation data from the spatially re-gridded, temporally homogenized multi-model ensemble of global climate simulation data through a statistical augmentation process comprising:
decomposing the temperature data into a first plurality of signal components, wherein a first signal component is related to El Nino-Southern Oscillation (ENSO), a second signal component is related to Atlantic Multidecadal Oscillation (AMO), a third signal component is related to Pacific Decadal Oscillation (PDO), and a fourth signal component is related to global warming;
generating a second plurality of surrogate time series from the first plurality of signal components, wherein at least one surrogate time series is generated by randomizing a phase of one of the first, second, and third signal components in a frequency domain;
generating a surrogate internal variability time series by resampling historical observational internal variability data; and
generating the synthetic simulation data based on the second plurality of surrogate time series, and the surrogate internal variability time series.
16. The system of claim 10 , wherein augmenting the spatially re-gridded, temporally homogenized multi-model ensemble of global climate simulation data comprises land-sea masking.
17. The system of claim 10 , wherein the predictive neural network is selected from the group consisting of a Convolutional Recurrent Neural Network (CRNN) having at least one Long Short-Term Memory (LSTM) layer, a Spherical Convolutional Neural Network (S 2 -CNN), and a transformer neural network.
18. The system of claim 10 , wherein the program code when executed by the processor further causes the processor to:
forecast at least one target climate variable at the target lead time using the NN-based climate forecasting model.
19. A non-transitory physical storage medium for training a neural network (NN)-based climate forecasting model, the storage medium comprising program code stored thereon, that when executed by a processor causes the processor to:
generate a multi-model ensemble of global climate simulation data by combining simulation data generated from at least two of a plurality of global climate models (GCMs);
determine a common spatial scale and a common temporal scale for the multi-model ensemble of global climate simulation data;
spatially re-grid the multi-model ensemble of global climate simulation data to the common spatial scale;
temporally homogenize the multi-model ensemble of global climate simulation data to the common temporal scale;
augment the spatially re-gridded, temporally homogenized multi-model ensemble of global climate simulation data with synthetic simulation data, to generate a spatially re-gridded, temporally homogenized, and augmented multi-model ensemble of global climate simulation data, wherein the synthetic simulation data is generated from the spatially re-gridded, temporally homogenized multi-model ensemble of global climate simulation data; and
train the NN-based climate forecasting model using the spatially re-gridded, temporally homogenized, and augmented multi-model ensemble of global climate simulation data, wherein the NN-based climate forecasting model comprises a predictive neural network, wherein for a given input into the NN-based climate forecasting model, a forecast output at a target lead time is generated by the NN-based climate forecasting model based on the given input, wherein the forecast output at the target lead time is compared to a corresponding desired output at the target lead time during the training, and wherein the given input and the corresponding desired output at the target lead time are selected from the multi-model ensemble of global climate simulation data.
20. The non-transitory physical storage medium of claim 19 , wherein the program code when executed by the processor further causes the processor to:
forecast at least one target climate variable at the target lead time using the NN-based climate forecasting model.