ADVANCEMENTS IN HYDROLOGICAL FORECASTING: IMPROVING ACCURACY AND OPERATIONAL EFFICIENCY THROUGH DEEP LEARNING MODELS

Abstract

Accurate and reliable hydrological forecasting models are essential for effective water resources management. A multitude of models have been developed for streamflow prediction, which can be divided into two categories: process-based and data-driven. Process-based models with one-, two- and three-dimensional simulation capabilities are based on physical principles, including state variables and theorized observable and scalable fluxes. Although models can provide insights into physical processes, they require large data sets, often come with enormous computational costs, and may entail considerable uncertainty. With breakthroughs in computational science in recent years, data-driven models armed with deep learning algorithms have become popular in time-series forecasting. These models are based on minimal historical data and do not require knowledge of underlying physical processes. Deep learning (DL) techniques are relatively less complex than physical models, making them easy to implement and providing satisfactory performance at low computational cost. The primary goal of this dissertation is to gain comprehensive knowledge of building deep learning approaches that can simultaneously improve accuracy, predictability, and computational efficiency in streamflow forecasting. This dissertation, introduces a series of advanced methods aimed at enhancing the accuracy and efficiency of ML models for hydrological forecasting tasks. These methods include: (i) a “correlation threshold” for partial autocorrelation and cross-correlation functions is proposed for input predictors selection; (ii) Wavelet transform couple with ML models for peak error reduction; (iii) Five optimization methods, Grid search (GS), Random search (RS), Bayesian optimization (BO), Particle swarm optimization (PSO), and Gray wolf optimization (GWO) have been built to optimize the hyperparameters of the ML models; (iv) Four model structures for short-term forecasting and long-term projection are systematized; (v) A novel dual-loop optimization framework has been developed to maximize ML models performance; (vi) The Transformer model has been proven effective for the multi-tasks forecasting problem. The comprehensive model structure, framework, and supporting methods proposed in this dissertation offer a promising solution for addressing both single-task and multi-task forecasting problems using machine learning models. This dissertation is expected to serve as practical reference material for highperformance ML-based hydrological prediction models.