Streamflow Forecast Errors and Their Impacts on Forecast-based Reservoir Flood Control

Abstract

To make full use of water captured by reservoirs in flood seasons, methods such as forecast-based reservoir flood control (i.e., reservoir operation during flood periods based on precipitation or inflow forecasts) have been developed in China in the past few decades. The success of forecast-based reservoir flood control depends heavily on the precision of the precipitation or inflow forecasts. This study analyzes the sources of uncertainty and quantifies it in the process of reservoir flood control based on forecasts. First, flood inflow series is regarded as a random process with the forecasted flood inflows as its expectation process; the flood inflow error series, which is the difference between the observed inflow process and the forecasted inflow process, is obtained according to the statistical characteristics of the flood forecasts. Second, reservoir flood routing simulations are carried out using the Monte Carlo method with the flood inflow error series as an input to obtain output water level and discharge series. Finally, the frequency distributions of the maximum water level (MWL) reached during the passage of floods are used to evaluate the uncertainties and the corresponding risks associated with forecast-based flood control. These uncertainties and risks are what the operators of the reservoirs care about the most.

A Novel Physics‐Aware Machine Learning‐Based Dynamic Error Correction Model for Improving Streamflow Forecast Accuracy

“…The time series plot of the error and observed discharge indicated the error sequence follows a definite pattern for the observed discharge (Figures 7g and 7h). Thus, error sequences are certainly demonstrating non‐Gaussian and time‐dependent behavior, and there exists a possibility of error modeling to improve the overall model performance (Cho & Kim, 2022; Farchi et al., 2021; Y. Li et al., 2014; Lipponen et al., 2013; Liu et al., 2015; McInerney et al., 2020; Nester et al., 2012; Shen, Ruijsch, et al., 2022; Shen, Wang, et al., 2022; Tajiki et al., 2020; Tian et al., 2014).…”

“…The time series plot of the error versus observed discharge was visually inspected to ensure the existence of any patterns (Cho & Kim, 2022; Shen, Ruijsch, et al., 2022; Shen, Wang, et al., 2022). Further, autocorrelation analysis of the error series is conducted to ensure error values at any time step are independent of the values at previous time steps, that is, error series has no patterns, that is, it is random and time‐independent (Lipponen et al., 2013; McInerney et al., 2020; Nester et al., 2012).…”

A Physics‐Aware Machine Learning‐Based Framework for Minimizing Prediction Uncertainty of Hydrological Models

“…Overall, it can be concluded that the error is non‐normal or non‐Gaussian. Thus, it is anticipated that further modelling of the error would improve the overall model performance resulting in an enhanced streamflow PI (Cho & Kim, 2022; Farchi et al., 2021; Lipponen et al., 2013; Liu et al., 2015; Y. Li et al., 2014; McInerney et al., 2020; Nester et al., 2012; Shen, Ruijsch, et al., 2022; Tajiki et al., 2020; Tian et al., 2014).…”

An Efficient Method to Correct Under-Dispersion in Ensemble Streamflow Prediction of Inflow Volumes for Reservoir Optimization

“…The alternative to adding variability to weather forcing is to add randomness to the model's initial conditions through ensemble data assimilation (Liu and Gupta 2007;Liu et al 2015). In snowmelt-dominated catchments, Snow Water Equivalent (SWE) is a major factor and can have considerable impact in volumetric seasonal forecasts (Clark and Hay 2004;Franz et al 2008;Zeinivand and De Smedt 2009;Sproles et al 2016).…”