Data-driven decadal climate forecasting using Wasserstein time-series generative adversarial networks

Recent trends in global climate modeling, coupled with the availability of more fine-scale datasets, have opened up opportunities for deep learning-based climate prediction to improve the accuracy of predictions over traditional physics-based models. For this, however, large ensembles of data are needed. Generative models have recently proven to be a suitable solution to this problem. For a sound generative model for time-series forecasting, it is essential that temporal dynamics are preserved in that the generated data obey the original data distributions over time. Existing forecasting methods aided by generative models are not adequate for capturing such temporal relationships. Recently, generative models have been proposed that generate realistic time-series data by exploiting the combinations of unsupervised and supervised learning. However, these models suffer from instable learning and mode collapse problems. To overcome these issues, here we propose Wasserstein Time-Series Generative Adversarial Network (WTGAN), a new forecasting model that effectively imitates the dynamics of the original data by generating realistic synthetic time-series data. To validate the proposed forecasting model, we evaluate it by backtesting the challenging decadal climate forecasting problem. We show that the proposed forecasting model outperforms state-of-the- art generative models. Another advantage of the proposed model is that once WTGAN is tuned, generating time-series data is very fast, whereas standard simulators consume considerable computer time. Thus, a large amount of climate data can be generated, which can substantially improve existing data-driven climate forecasting models.