Abstract

Hydrological time-series forecasting and analysis play a crucial role to expand knowledge and insight into hydrological processes as well as for water resources management agencies in making decisions and policies in the monitor, prevent and mitigate losses caused by natural disasters such as floods, droughts, or climate change. However, the instability, uncertainty, or sometimes inadequacy of hydrological time-series data are common problems faced, due to characteristics of hydrological systems inherently heterogeneous. The model-driven approach has a long tradition that is widely used around the world as a trustworthy approach for studying extreme events, simulating, and forecasting hydrological processes. These models have a tendency to require a large number of various types of data that may not always be available or could be difficult to obtain. Moreover, various disadvantages have been identified in a variety of previous studies, such as performance or weakness in uncertainty analysis.

On the other hand, the data-driven approach, especially deep learning neural network models, has received an increasing interest of scientists in the analysis and forecasting of hydrological data in recent years, due to minimizing the assumptions of physical processes in constructing a model. Besides, deep learning models have demonstrated superiority in a myriad of complex tasks, such as natural language processing, object identification, and face recognition. Despite remarkable results, the applications of deep learning neural networks in hydrology are modest compared to other disciplines. There is a gap in the study and application of advances of deep learning algorithms in the hydrological issues. This thesis focuses on exploiting the latest advances of deep learning algorithms as well as the deep learning neural networks in the application, analysis, and forecasting of hydrological time-series data. Main focuses of this thesis are as follows:

First, several special architectures of recurrent neural network (RNN) models, such as long short-term memory (LSTM) neural network and the gated recurrent unit (GRU) neural network, are designed to address the hydrological forecast problem for the short-term periods (e.g. hourly, 6-hourly, or daily), in particular, real-time streamflow prediction, flood forecasting, and tidal river water level forecasting. A common characteristic of these problems is the use of point-data series (or 1-dimensional sequential data) as input data for deep learning models. These data series can be data on flowrate or water level which were observed at hydro-meteorological stations upstream or downstream of the study area.

Next, several deep learning models based on convolution neural networks (CNNs) are constructed to solve the bias correction problem of gridded satellite-based precipitation products. The structure of the proposed models could be a combination of the CNN model and several various architectural paradigms that belong to the supervised learning class according to the classification of the deep learning algorithms, for example, autoencoder or LSTM architectures. In addition, grid-data series (or 2-dimensional data) are used as input data in these CNN-based models.

The thesis has investigated, examined, and applied the state-of-the-art advances of deep learning neural networks in hydrological time-series forecasting and analysis. The findings of this thesis highlighted the potential in the application of deep learning neural network models in studies on hydrology and water resources.