Hydrological probabilistic forecasting based on deep learning and Bayesian optimization algorithm

Haijun Bai,Zhendong Zhang,Bin Li,Hui Qin,Changming Liu,Guanjun Li

doi:10.2166/nh.2021.161

Haijun Bai, Zhendong Zhang + Show 4 more

Open Access

https://doi.org/10.2166/nh.2021.161

Copy DOI

Abstract

Abstract Obtaining accurate runoff prediction results and quantifying the uncertainty of the forecasting are critical to the planning and management of water resources. However, the strong randomness of runoff makes it difficult to predict. In this study, a hybrid model based on XGBoost (XGB) and Gaussian process regression (GPR) with Bayesian optimization algorithm (BOA) is proposed for runoff probabilistic forecasting. XGB is first used to obtain point prediction results, which can guarantee the accuracy of forecast. Then, GPR is constructed to obtain runoff probability prediction results. To make the model show better performance, the hyper-parameters of the model are optimized by BOA. Finally, the proposed hybrid model XGB-GPR-BOA is applied to four runoff prediction cases in the Yangtze River Basin, China and compared with eight state-of-the-art runoff prediction methods from three aspects: point prediction accuracy, interval prediction suitability and probability prediction comprehensive performance. The experimental results show that the proposed model can obtain high-precision point prediction, appropriate prediction interval and reliable probabilistic prediction results on the runoff prediction problems.

Highlights

With the increasing shortage of energy, hydropower has received worldwide attention as a clean renewable energy (Liu et al 2018a)
Experimental design and parameter settings To verify the performance of the proposed method, LGB (Deng et al 2018), Gradient Boosting Regression Tree (GBR) (Rao et al 2019), long short-term memory (LSTM) network (Zhang et al 2019a), convolutional neural network (CNN) (Le Callet et al 2006), Artificial neural network (ANN) (Tan et al 2018), support vector regression (SVR) (Luo et al 2019), Quantile regression (QR) (Zhang et al 2019b) and Gaussian process regression (GPR) (Sun et al 2014) are compared with XGB
Using the framework proposed in this paper, combined with GPR, these models are further transformed to obtain the probabilistic forecasting model: XGB-GPR, LGBGPR, GBR-GPR, LSTM-GPR, CNN-GPR, ANN-GPR and SVR-GPR

Summary

Introduction

With the increasing shortage of energy, hydropower has received worldwide attention as a clean renewable energy (Liu et al 2018a). The process-driven model is based on the hydrological concept and focuses on the description of the physical mechanism of runoff yield and concentration, such as the Xin’anjiang hydrological model (Fang et al 2017) and numerical weather prediction (Wu & Lin 2017) These models have high prediction accuracy and are highly interpretable, but their data collection is difficult and the solution is time-consuming (Wu & Lin 2017). The time-series model is a commonly used method for runoff prediction, mainly including auto-regressive model (AR), moving average model (MA), auto-regressive moving average model (ARMA) and their variants (Papacharalampous et al 2018) These models are based on data stationarity assumptions, so their prediction accuracy is limited because of the strong nonlinearity of runoff (Mauricio 1995). Artificial neural network (ANN) is used to characterize the nonlinearity of

Methods

Results

Discussion

Conclusion