Runoff Forecasting Using Machine-Learning Methods: Case Study in the Middle Reaches of Xijiang River.

Lu Xiao,Dawei Zha,Ming Zhong

doi:10.3389/fdata.2021.752406

Abstract

Runoff forecasting is useful for flood early warning and water resource management. In this study, backpropagation (BP) neural network, generalized regression neural network (GRNN), extreme learning machine (ELM), and wavelet neural network (WNN) models were employed, and a high-accuracy runoff forecasting model was developed at Wuzhou station in the middle reaches of Xijiang River. The GRNN model was selected as the optimal runoff forecasting model and was also used to predict the streamflow and water level by considering the flood propagation time. Results show that (1) the GRNN presents the best performance in the 7-day lead time of streamflow; (2) the WNN model shows the highest accuracy in the 7-day lead time of water level; (3) the GRNN model performs well in runoff forecasting by considering flood propagation time, increasing the Qualification Rate (QR) of mean streamflow and water level forecast to 98.36 and 82.74%, respectively, and illustrates scientifically of the peak underestimation in streamflow and water level. This research proposes a high-accuracy runoff forecasting model using machine learning, which would improve the early warning capabilities of floods and droughts, the results also lay an important foundation for the mid-long-term runoff forecasting.

Highlights

Runoff forecasting is the foundation of water resource management, deployment, and efficient utilization
The results show that the mean streamflow, water level, and meteorological factors of Wuzhou station before 7, 10, and 15 days are significantly correlated with the mean streamflow and water level of the day
Taking the forecast results by BP neural network as an example, the Mean absolute error (MAE) values of mean streamflow for 7, 10, and 15-day lead time are 1,772.7856, 1,934.0324, and 2,098.2541 m3·s−1 respectively; Deterministic coefficient (DC) values are 0.2081, 0.0951 and −0.2322, respectively; R2 values are 0.5224, 0.4541, and 0.4333, respectively; Mean relative error (MRE) values are 0.2630, 0.2995, and 0.3715, respectively; Root mean square error (RMSE) values are 3,036.2640, 3,268.3675, and 3,304.2589 m3·s−1 respectively; Qualification Rate (QR) values are 64.88, 68.86, and 53.50%, respectively

Summary

Introduction

Runoff forecasting is the foundation of water resource management, deployment, and efficient utilization. It is of great significance to reservoir operation, water resource emergency scheduling, hydro-power generation, and irrigation management decisions (Niu et al, 2018). The river runoff is sensitive to various factors, such as catchment response times and the accuracy of meteorological forecasts with time variability and uncertainty (Lima et al, 2016). The establishment of hydrological models provides important support for runoff forecasting. The runoff process is simulated and forecasted from the perspective of the physical mechanism. Hydrological model driving relies on the input of a large amount of meteorological data and watershed characteristics parameters. The forecasting process is relatively complicated, and its accuracy is limited to the accuracy and completeness of the data (Nourani, 2017)

Objectives

Methods

Results

Conclusion