Integrating Novel Hybrid Approaches to Handle Complex Hydrological Time Series Data: A Case Study on Rivers Inflow in Indus Basin System

Abstract

The hydrological time series data is a crucial area of scientific study which becomes a subject of great interest now a day. Prediction of hydrological time series data is vital for the planning, management and optimal allocation of water resources. However, accurate prediction of hydrological time series data becomes a challenging task due to abrupt changes in rainfall, temperature, evaporations, and extreme climate conditions which makes the hydrological data non-linear, non-stationary corrupted with time-varying and nosiest characteristics. During the past decades, several traditional time series and data-driven models have been proposed to predict and analyze the hydrological time series data. But these traditional time series and datadriven models are only useful for linear and non-linear problems respectively. Both approaches did not consider the time-varying and nosiest characteristics of complex hydrological data which reduces their prediction efficiency. To overcome these shortcomings, effective data preprocessing methods have been introduced in literature to eliminate noises and extracting multiscale characteristics which enhanced the prediction performance of traditional time series and data-driven models. In this regard, this thesis develops various hybrid models in which different advanced signal and statistical data pre-processing methods have been adapted to denoise and decompose the complex hydrological time series data which have a significant impact on the performance of traditional statistical and data-driven models as irregular and non-linear inputs often lead to wrong outputs. All the proposed hybrid models comprised on ‘denoised, decomposed-threshold, prediction and aggregation’ to model hydrological time series data. Our proposed models shown that by effectively removing noises and reducing complexities of hydrological time series data through extracting multi-scale components, existing models can provide more accurate results rather than using direct data. The efficiency of proposed hybrid models is illustrated by using four rivers inflow data of the Indus Basin System. In this perspective, five hybrid models have been proposed in this thesis with aim to increase the prediction of complex hydrological time series data. First four hybrid models are proposed only for single-site rivers inflow data i.e. univariate time series and one hybrid model is proposed for multi-site rivers inflow data i.e. multivariate time series. The layout of first four proposed methodology is same for all univariate time series models as; first, noises have been removed with appropriate methods; second, denoised series have been decomposed in different Intrinsic ii Mode Decomposition Functions (IMF). The denoised IMFs which proved to be non-linear are further threshold so that sparsities can be removed efficiently. Finally all IMFs are predicted through simple statistical and data driven-models to get final prediction. To model the multi-site rivers inflow data, application of Vine copula has been combined with previously proposed hybrid model where both uncorrelated and correlated assumptions of multisite rivers inflow data have been considered. For the first assumption, independent residuals have been generated from previously proposed models and by incorporating the second assumption i.e. dependence structure between multi-site rivers inflow, prediction performance of individual rivers inflow is enhanced. Our proposed models showed that by appropriately considering the nature of complex hydrological time series data or by adding additional information regarding to data, more accurate prediction results can be obtained.