GRU-Transformer洪水预报模型构建与应用  

Construction and Application of GRU-Transformer Flood Forecasting Model

作　　者：李文忠 刘成帅 邬强 胡彩虹[1] 解添宁 田露 LI Wen-zhong;LIU Cheng-shuai;WU Qiang;HU Cai-hong;Xie Tian-ning;TIAN Lu(Yellow River Laboratory,Zhengzhou University,Zhengzhou 450001,Henan Province,China)

机构地区：[1]郑州大学黄河实验室,河南郑州450001

出　　处：《中国农村水利水电》2023年第11期35-44,共10页China Rural Water and Hydropower

基　　金：国家自然科学基金项目(U2243219、51979250);黄河实验室(郑州大学)一流课题专项基金(YRL22IR02);河南省科技攻关(222102320455)。

摘　　要：洪水预报是黄河中游地区防洪减灾重要的非工程措施之一。研究通过耦合门控制循环单元(Gate Recurrent Unit,GRU)和Transformer机器学习模型,构建GRU-Transformer洪水预报模型,对黄河中游地区典型流域进行降雨径流模拟预测。同时将预报结果与ANN(Artificial Neural Network)、WOA-GRU(Whale Optimization Algorithm Gate Recurrent Unit)神经网络洪水预报模型的预报结果对比分析,着重探索如何将Transformer模型更好地应用于洪水预报领域,旨在提高黄河中游地区洪水预报精度。采用故县水库控制流域1990-2016年49场历史观测洪水数据,以24个站点实测降雨量及出口断面流量作为输入,不同预见期下的洪水过程作为输出,率定期为39场次,验证期为10场次。研究结果表明:GRU-Transformer模型在洪水预报种具有较好的适用性,在预见期1~6 h洪水预报中,GRU-Transformer预报精度较高,校准期和验证期的NSE均大于0.85,且预报精度在相同预见期下优于WOA-GRU和ANN模型,但预报精度会随着预见期增大而出现一定程度的下降;GRU-Transformer模型较稳定地更好预测洪峰,且在较小流量洪水过程预报时及洪水退水阶段模拟效果表现出优异效果,但随预见期增加出现低估洪峰现象;GRU-Transformer模型比WOA-GRU和ANN模型具有更好的鲁棒性,随洪水预见期增大,其预报精度呈缓慢下降,降低的最慢。因此,GRU-Transformer模型可以作为较好的洪水预报方法之一,相关成果为流域防洪安全提供了新的预报方法及科学决策依据。Flood forecasting is one of the important non-engineering measures for flood control and disaster reduction in the middle reaches of the Yellow River.In this study,a GRU-Transformer flood forecasting model is constructed by coupling gated recurrent units(GRU)with Transformer machine learning models,and rainfall-runoff simulations are conducted to predict flood events in typical sub-basins of the mid⁃dle reaches of the Yellow River.The predictive results are compared and analyzed with those of the ANN(Artificial Neural Network)and WOA-GRU(Whale Optimization Algorithm Gate Recurrent Unit)neural network flood forecasting models,with a focus on exploring how to better apply the Transformer model to the field of flood forecasting in order to improve the accuracy of flood forecasting in the middle reaches of the Yellow River.The model is established by using historical observed flood data from 1990 to 2016 in the Gu County Reservoir Con⁃trolled Basin.The input data includes rainfall data measured at 24 stations and discharge data at the outlet cross-section,while the output da⁃ta includes flood events under different lead times.The model is calibrated by using 39 flood cases and validated using 10 flood cases.The re⁃sults of the study show that the GRU-Transformer model has good applicability in flood forecasting,exhibiting higher predictive accuracy in the 1~6 hour lead time flood forecasting,with NSE values of greater than 0.85 in both calibration and validation periods.Its predictive accu⁃racy is better than the WOA-GRU and ANN models under the same lead time,but decreases to a certain extent with increasing lead times.The GRU-Transformer model is more stable and better at predicting flood peaks,showing excellent performance in predicting small flow flood processes and simulating the recession phase of floods.However,it tendes to underestimate flood peaks with longer lead times.Com⁃pared with the WOA-GRU and ANN models,the GRU-Transformer model has better robustness,and its predictive accuracy decreases

关 键 词：降雨径流模拟 洪水预报 神经网络 GRU-Transformer 黄河中游 

分 类 号：TU992[建筑科学—市政工程]