Development of generative adversarial networks for spatiotemporal fluid flow, atmospheric and flood predictions

Abstract

Accurate prediction of fluid flow is of vital importance to many applications in engineering and physics. Long-standing challenges in flow simulations are the simultaneous high accuracy and efficiency demanded in numerical computations. Recent advances in machine learning technologies are increasingly of interest for the efficient simulation of nonlinear and complex systems. Deep learning techniques are capable of capturing the physical dynamics without prior knowledge of underlying physical relationships.

However, there remain several challenges in nonlinear fluid flow modelling, i.e., spatiotemporal nonlinear fluid flow modelling, large data-driven fluid flow modelling, real-time forecasting of nonlinear fluid flows beyond the training period, improvement of the forecasting accuracy in long lead-time. To overcome these challenges, in this work, Generative Adversarial Networks (GANs) have been first introduced to spatiotemporal nonlinear fluid flow prediction while data assimilation techniques are used for improving the accuracy of long term forecasting.

The key contributions of this thesis are: firstly, an Artificial Intelligence (AI) fluid model based on a deep convolutional generative adversarial network (DCGAN) has been developed for modelling spatiotemporal flow distributions. Secondly, a hybrid deep adversarial autoencoder (VAE-GAN) model to integrate generative adversarial network (GAN) and variational autoencoder (VAE) has been proposed for large data-driven nonlinear fluid flow prediction. Thirdly, a real-time predictive machine learning model, i.e., artificial neural network (ANN), long short term memory (LSTM), DCGAN, and VAE-GAN has been developed for basin streamflow, urban flooding, and national air pollution forecasting problems respectively. Fourthly, an ensemble Kalman filter (EnKF) for GAN and convolutional LSTM (GAN-ConvLSTM) based forecasting system has been proposed for accurate long lead-time forecasting.

The presented models have been validated by various test cases and the results are in good agreement with high-fidelity models and observations. Promising results have shown that the DCGAN and VAE-GAN models are capable of accurately predicting the spatiotemporal flow features as the flow evolves, with CPU speed-up of several orders of magnitude. In addition, the ANN, LSTM, DCGAN and VAE-GAN models have demonstrated to provide efficient and accurate forecasts for a long lead-time in applications of streamflow, flooding and ozone forecasting in spatial and temporal spaces. Finally, the two-hybrid forecast models (DCGAN-EnKF and ConvLSTM-EnKF) are able to yield long lead-time forecasts of dynamic states and the use of EnKF in ConvLSTM and DCGAN models successfully corrects online model errors and significantly improves the real-time forecasting of dynamic systems.