Modeling and simulating of reservoir operation using the artificial neural network, support vector regression, deep learning algorithm

Di Zhang,Junqiang Lin,Qidong Peng,Dongsheng Wang,Tiantian Yang,Soroosh Sorooshian,Xuefei Liu,Jiangbo Zhuang

doi:10.1016/j.jhydrol.2018.08.050

Abstract

Reservoirs and dams are vital human-built infrastructures that play essential roles in flood control, hydroelectric power generation, water supply, navigation, and other functions. The realization of those functions requires efficient reservoir operation, and the effective controls on the outflow from a reservoir or dam. Over the last decade, artificial intelligence (AI) techniques have become increasingly popular in the field of streamflow forecasts, reservoir operation planning and scheduling approaches. In this study, three AI models, namely, the backpropagation (BP) neural network, support vector regression (SVR) technique, and long short-term memory (LSTM) model, are employed to simulate reservoir operation at monthly, daily, and hourly time scales, using approximately 30 years of historical reservoir operation records. This study aims to summarize the influence of the parameter settings on model performance and to explore the applicability of the LSTM model to reservoir operation simulation. The results show the following: (1) for the BP neural network and LSTM model, the effects of the number of maximum iterations on model performance should be prioritized; for the SVR model, the simulation performance is directly related to the selection of the kernel function, and sigmoid and RBF kernel functions should be prioritized; (2) the BP neural network and SVR are suitable for the model to learn the operation rules of a reservoir from a small amount of data; and (3) the LSTM model is able to effectively reduce the time consumption and memory storage required by other AI models, and demonstrate good capability in simulating low-flow conditions and the outflow curve for the peak operation period.

Highlights

Half of the major global river systems are affected by reservoirs and dams, and human beings manage and utilize water resources through reservoirs for power generation, water supply, navigation, disaster prevention, flood control and mitigation, drought relief (Dynesius and Nilsso, 1994; WCD, 2000; ICOLD, 2011; Lehner et al, 2011; Shang et al, 2018)
The observed GZB reservoir outflows are compared with simulated results based on the three different artificial intelligence (AI) models, i.e., the BP neural network, support vector regression (SVR), and long short-term memory (LSTM) model, combined with various model parameters
At monthly scale, the simulation accuracy is ranked as Sigmoid > Polynomial > Radial Basis Function (RBF) > Linear among the different kernel functions, at daily scale, the best accuracy ranking is Sigmoid > RBF > Linear > Polynomial; and at hourly scale, the best accuracy ranking is Sigmoid > RBF > Polynomial > Linear (Table 4)

Summary

Introduction

Half of the major global river systems are affected by reservoirs and dams, and human beings manage and utilize water resources through reservoirs for power generation, water supply, navigation, disaster prevention, flood control and mitigation, drought relief (Dynesius and Nilsso, 1994; WCD, 2000; ICOLD, 2011; Lehner et al, 2011; Shang et al, 2018). Starting in the 1980s, with the development of hydrology, hydraulics and river dynamics, conceptual or physical-based models (such as HEC-ResSim, WEAP21, etc.) have been proposed and are widely used in reservoir hydrological process simulation and reservoir operation decisions (Klipsch and Hurst, 2003; Yates et al, 2005). Such models transform the empirical, mechanical, and blind operation patterns of early reservoir operations that were based on historical hydrological statistics, operated by so-called rule curves. Physical-based models provide a more practical physical and mathematical basis for the calculation of controlled releases or storage (See Table 1)

Objectives

Methods

Results

Discussion

Conclusion