Development of Deep Learning Hybrid Models for Hydrological Predictions

Development of Deep Learning Hybrid Models for Hydrological Predictions

Forecasting hydrologic phenomena are critical for strategic environmental planning, designing the hydrologic structures, and managing agricultural practices and water resources. Physical models are the mainstream method that helps understand the physical mechanisms and dynamics used in hydrological predictions. They are used for addressing the characteristics of hydrological phenomena while considering the initial conditions and spatial-temporal resolution of the model inputs. Data-driven models, on the other hand, are based on artificial intelligence and are designed as alternatives to discover the relationships between a set of predictors and a target variable without considering any of the initial conditions or underlying assumptions. These methods are relatively new and are becoming state-of-the-art to address different prediction problems.

This doctoral thesis, with its five primary objectives, aims to build a set of deep learning hybrid models and evaluate for their predictive skills in forecasting hydrological variables such as soil moisture (SM), evapotranspiration (ETo), and streamflow water levels (SWL) within Australian Murray-Darling Basin. The first objective establishes the significance of feature selection to predict the monthly SWL at six study sites. The BRF-LSTM hybrid method integrated with the long-short term memory (LSTM) model with a Boruta-Random forest optimizer (BRF) is used to demonstrate the importance of feature selection for SWL forecasting problems. The second objective is to develop a CNN-GRU hybrid model using the ant colony optimization to screen the most correlated features from a diversified set of inputs using convolutional neural network (CNN) and gated recurrent unit (GRU) networks for evapotranspiration (ETo) forecasting. The results show that the CNNGRU model integrated with the ACO method has outperformed the benchmark models over multi-step forecast horizons, and it has also captured the complex and non-linear relationships between predictors and daily ETo. The third objective employs the BRF feature selection method to identify the global climate model (GCM)-simulated variables for an LSTM model, aiming to estimate upper-layer surface soil moisture (SM) under RCP4.5 and 8.5 warming scenarios. The results demonstrate that the proposed BRF-LSTM model is more accurate than benchmark models, and this objective has established a new approach that can deal with GCM-simulated variables.

The fourth objective develops a CEEMDAN-CNN-GRU hybrid model to forecast daily surface soil moisture (SSM) by using neighbourhood component analysis (NCA), complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN), convolutional neural networks (CNN), and gated recurrent units (GRU). The CEEMDAN-CNN-GRU hybrid model outperforms all benchmark and standalone models in simulating surface soil moisture. The fifth objective is to develop the CBILSTM hybrid model, coupled with CEEMDAN and a variational mode decomposition (VMD) to build the CVMD-CBiLSTM hybrid model for streamflow water level forecasting. This proposed model reveals that the CVMD-CBiLSTM hybrid model had outperformed the benchmark models.

The artificial intelligence (AI) methodologies developed in this PhD project are expected to be a significant step forward in developing AI-based data-driven decision support systems that will enable hydrologists and climate specialists to design water resource management strategies. Though this work focuses on soil moisture, evapotranspiration, and streamflow water level forecasting, the developed methodologies can also contribute significantly to other areas, such as flood forecasting, irrigation scheduling, and sustainable management of water resources. Overall, the doctoral study establishes significant scientific pathways for water resources management and smart farming using AI-based decision support systems.

File reproduced in accordance with the copyright policy of the publisher/author.