Methods and systems for climate forecasting using artificial neural networks

1 . A method for generating a neural network (NN)-based climate forecasting model for a target climate forecast application, comprising:

selecting a first global climate simulation dataset from a plurality of candidate simulation datasets, wherein each candidate simulation dataset is generated from a different global climate simulation model (GCM);

training the NN-based climate forecasting model on the selected first global climate simulation dataset, wherein the NN-based climate forecasting model comprises a first predictive neural network, and wherein each input and corresponding desired output used in the training are elements of the selected first global climate simulation dataset;

tuning the NN-based climate forecasting model on reanalysis data by freezing one or more layers of the predictive neural network and training the NN-based climate forecasting model on the reanalysis data, wherein the reanalysis data is generated from assimilating a first set of observational historical data with one or more physical dynamical models;

validating the NN-based climate forecasting model on a second set of observational historical climate data to generate a first validation result;

in response to the first validation result, selecting a second global climate simulation dataset from the plurality of candidate simulation datasets;

training the NN-based climate forecasting model on the selected second global climate simulation dataset, wherein each input and corresponding desired output used in the training are elements of the selected second global climate simulation dataset;

validating the NN-based climate forecasting model on the second set of observational historical climate data to generate a second validation result;

in response to the second validation result, forecasting a primary target climate variable at a target lead time using the NN-based climate forecasting model;

forecasting a secondary climate variable at a plurality of geographical locations using a second predictive neural network, wherein an input of the second predictive neural network is connected to the output of the NN-based climate forecasting model to intake the primary target climate variable; and

generating for display, a map showing the secondary climate variable at the plurality of geographical locations.

2 . The method of claim 1 , wherein the predictive neural network is a Convolutional Recurrent Neural Network (CRNN) having at least one Long Short-Term Memory (LSTM) layer.

3 . The method of claim 1 , wherein the predictive neural network is a Spherical Convolutional Neural Network (S 2 -CNN).

4 . The method of claim 1 , wherein the selecting the first global climate simulation dataset comprises:

validating each of the plurality of candidate simulation datasets, by comparing a sample statistic of at least one climate variable, between simulation data from each candidate simulation dataset and a third set of observational historical climate data, wherein the at least one climate variable is selected based on the target climate forecast application; and

computing a forecast skill score for each validated candidate simulation dataset.

5 . The method of claim 1 , further comprising:

augmenting each candidate simulation dataset using at least one of climatology augmentation, temporal augmentation, and McKinnon augmentation.

6 . The method of claim 1 , wherein each global climate simulation model is selected from the group consisting of CNRM-CM5 model, MPI-ESM-LR model, GISS-E2-H model, NorESM1-M model, HadGEM2-ES model, and GFDL-ESM2G model.

7 . A system for generating 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:

select a first global climate simulation dataset from a plurality of candidate simulation datasets, wherein each candidate simulation dataset is generated from a different global climate simulation model (GCM);

train the NN-based climate forecasting model on the selected global climate simulation dataset, wherein the NN-based climate forecasting model comprises a predictive neural network, and wherein each input and corresponding desired output used to train the NN-based climate forecasting model are elements of the selected first global climate simulation dataset;

tune the NN-based climate forecasting model on reanalysis data by freezing one or more layers of the predictive neural network and training the NN-based climate forecasting model on the reanalysis data, wherein the reanalysis data is generated from assimilating a first set of observational historical data with one or more physical dynamical models;

validate the NN-based climate forecasting model using a second set of observational historical climate data to generate a first validation result;

in response to the first validation result, select a second global climate simulation dataset from the plurality of candidate simulation datasets;

train the NN-based climate forecasting model on the selected second global climate simulation dataset, wherein each input and corresponding desired output used in the training are elements of the selected second global climate simulation dataset;

validate the NN-based climate forecasting model on the second set of observational historical climate data to generate a second validation result;

in response to the second validation result, forecast a primary target climate variable at a target lead time using the NN-based climate forecasting model;

forecast a secondary climate variable at a plurality of geographical locations using a second predictive neural network, wherein an input of the second predictive neural network is connected to the output of the NN-based climate forecasting model to intake the primary target climate variable; and

generate for display, a map showing the secondary climate variable at the plurality of geographical locations.

8 . The system of claim 7 , wherein the predictive neural network is a Convolutional Recurrent Neural Network (CRNN) having at least one Long Short-Term Memory (LSTM) layer.

9 . The system of claim 7 , wherein the predictive neural network is a Spherical Convolutional Neural Network (S 2 -CNN).

10 . The system of claim 7 , wherein the program code to select the first global climate simulation dataset, when executed by the processor, causes the processor to:

validate each of the plurality of candidate simulation datasets, by comparing a sample statistic of at least one climate variable, between simulation data from each candidate simulation dataset and a third set of observational historical climate data, wherein the at least one climate variable is selected based on the target climate forecast application; and

compute a forecast skill score for each validated candidate simulation dataset.

11 . The system of claim 7 , wherein the program code, when executed by the processor, further causes the processor to:

augment each candidate simulation dataset using at least one of climatology augmentation, temporal augmentation, and McKinnon augmentation.

12 . The system of claim 7 , wherein each different global climate simulation model is selected from the group consisting of CNRM-CM5 model, MPI-ESM-LR model, GISS-E2-H model, NorESM1-M model, HadGEM2-ES model, and GFDL-ESM2G model.

13 . A non-transitory physical storage medium for generating 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:

select a first global climate simulation dataset from a plurality of candidate simulation datasets, wherein each candidate simulation dataset is generated from a different global climate simulation model (GCM);

train the NN-based climate forecasting model on the selected global climate simulation dataset, wherein the NN-based climate forecasting model comprises a predictive neural network, and wherein each input and corresponding desired output used to train the NN-based climate forecasting model are elements of the selected first global climate simulation dataset;

tune the NN-based climate forecasting model on reanalysis data by freezing one or more layers of the predictive neural network and training the NN-based climate forecasting model on the reanalysis data, wherein the reanalysis data is generated from assimilating a first set of observational historical data with one or more physical dynamical models;

validate the NN-based climate forecasting model using a second set of observational historical climate data

in response to the first validation result, select a second global climate simulation dataset from the plurality of candidate simulation datasets;

train the NN-based climate forecasting model on the selected second global climate simulation dataset, wherein each input and corresponding desired output used in the training are elements of the selected second global climate simulation dataset;

validate the NN-based climate forecasting model on the second set of observational historical climate data to generate a second validation result;

in response to the second validation result, forecast a primary target climate variable at a target lead time using the NN-based climate forecasting model;

forecast a secondary climate variable at a plurality of geographical locations using a second predictive neural network, wherein an input of the second predictive neural network is connected to the output of the NN-based climate forecasting model to intake the primary target climate variable; and

generate for display, a map showing the secondary climate variable at the plurality of geographical locations.

14 . The non-transitory physical storage medium of claim 13 , wherein the predictive neural network is a Convolutional Recurrent Neural Network (CRNN) having at least one Long Short-Term Memory (LSTM) layer.