Comparing the Performance of an Autoregressive State-Space Approach to the Linear Regression and Artificial Neural Network for Streamflow Estimation

Abstract

Accurate streamflow estimation remains a great challenge although diverse modeling techniques have been developed during recent decades. In contrast to the process-based models, the empirical data-driven methods are easy to operate, require low computing capacity and yield fairly accurate outcomes, among which the state-space (STATE) approach takes use of the temporal structures inherent in streamflow series and serves as a feasible solution for streamflow estimation. Yet this method has rarely been applied, neither its comparison with other methods. The objective was to compare the performance of an autoregressive STATE approach to the traditional multiple linear regression and artificial neural network in simulating annual streamflow series of 15 catchments located in the Loess Plateau of China. Annual data of streamflow (Q), precipitation (P) and potential evapotranspiration (PET) during 1961 ~ 2013 were collected. The results show that STATE was generally the most accurate method for Q estimation, explaining almost 90% of the total variance averaged over all the 15 catchments. The estimation of streamflow relied on its own of the previous year for most catchments. Besides, the impacts of P and PET on the temporal distribution of streamflow were almost equal. Missing data were estimated using the STATE method, which allowed inter-annual trend analysis of the streamflow. Significant downward trends were manifested at all the 15 catchments during the study period and the corresponding slopes ranged from -0.24 to -1.71 mm y -1 . These findings hold important implications for hydrological modelling and management in China's Loess Plateau and other arid and semi-arid regions.

A Factorial Ecological-Extended Physical Input-Output Model for Identifying Optimal Urban Solid Waste Path in Fujian Province, China

“…In fact, a USWS has complexities related to different production technologies, industry scales, and pollution intensities. Valuable information is often hidden under the interrelationships between these factors and the consequent effects [20,21]. For example, variations in metal productive capability can affect the amount of solid waste delivered to the electrical equipment manufacture sector, as well as the amount of solid waste received from the metal ore mining sector.…”

A Stepwise Clustered Hydrological Model for Addressing the Temporal Autocorrelation of Daily Streamflows in Irrigated Watersheds

“…They are flexible enough to approximate various complex processes and interrelationships, and allow for direct mapping from meteorological and ancillary inputs (e.g., soil moisture (Schmidt et al., 2020), catchment attributes (Kratzert, Klotz, Shalev, et al., 2019) and irrigation scheduling (Mohan & Vijayalakshmi, 2009)) to streamflow or other output fluxes (Solomatine & Ostfeld, 2008). Many data‐driven hydrological models have been developed (Adnan et al., 2020; ASCE Task Committee, 2000a; Y. Yang et al., 2020; H. Zhang et al., 2019). However, most of the existing modeling efforts are focused on short‐term predictions, such as single‐ or multi‐step‐ahead forecasting (Adnan et al., 2019; Badrzadeh et al., 2013; Bray & Han, 2004; Campolo et al., 2003; Fleming et al., 2015; He et al., 2014; Kisi et al., 2012; Toth et al., 2000; J. Yang et al., 2013).…”

A Classification-Based Machine Learning Approach to the Prediction of Cyanobacterial Blooms in Chilgok Weir, South Korea

“…Numerous earlier studies attempted to predict the cyanobacterial blooms accurately by developing process-based models that mathematically provide the mechanism of the blooms [16]. Nevertheless, the process-based models require considerable input [16] and computing time [18] as they all involve related factors such as water quality, climate, and flow rate. On the other hand, data-driven models using machine learning or deep learning produce output by taking less running time [18] and only some main factors [16].…”

“…Nevertheless, the process-based models require considerable input [16] and computing time [18] as they all involve related factors such as water quality, climate, and flow rate. On the other hand, data-driven models using machine learning or deep learning produce output by taking less running time [18] and only some main factors [16]. Some research proved that the data-driven models employing techniques such as Random Forest (RF) [19,20], Support Vector Machine (SVM) [19], ANN [19], and Extreme Learning Machine (ELM) [21] ensured high accuracy in predicting the real-valued output such as cyanobacterial cell density [20] or Chlorophyll-a concentration (Chl-a), which is a proxy index for the cyanobacterial blooms [17,20,21].…”