洪水预报区域化LSTM模型研究

Research on LSTM-based Regionalized Flood Forecasting Model

受水文气象资料条件限制,在乏资料地区开展洪水预报仍面临挑战。虽然水文模型参数区域化是解决该问题的常用途径,但机器学习模型因其具有建模简单、使用方便等特点,已逐渐成为构建乏资料地区洪水预报方案的重要方法。以山东省南四湖湖西平原区为例,借鉴水文地区综合的思想,对2010~2021年8个流域共40场暴雨洪水资料进行合成,构建了基于长短期记忆神经网络(LSTM)的区域化洪水预报模型。研究结果表明,区域化LSTM洪水预报模型能够较好地模拟实际洪水过程,训练集和测试集的洪峰流量相对误差均小于10%,纳什效率系数均大于0.9;在15 h预见期内,区域化洪水预报模型具有较高的预报精度,当预见期超过15 h时,模型的预报精度有所降低。

Limited by hydrometeorological data,flood forecasting in ungauged basins still faces challenges.Parameter regionalization is a common method to solve this problem.The machine learning model has the characteristics of simple modeling and convenient use compared with the traditional flood forecasting model.Taking the West Plain of Nansihu Lake in Shandong Province as the research area,referencing the idea of hydrological regional synthesis,this paper synthesizes the data of 40 floods in 8 watersheds from 2010 to 2021,and builds a regionalized flood forecasting model based on Long Short-Term Memory(LSTM).The results show that the regionalized flood forecasting model can simulate the actual flood process well,the relative error of flood peak in both the training set and the testing set are less than 10%,and the Nash-Sutcliffe efficiency coefficients are all greater than 0.9;In the 15 h forecast period,the regionalized flood forecasting model has higher forecasting accuracy,and when the forecast period is more than 15 h,the forecast accuracy of the model decreases.