基于LSTM、RF、SVR三种机器学习方法的径流预测研究

Research on Runoff Prediction Based on Machine Learning:LSTM,RF and SVR

为探究不同预报方案对机器学习模型径流预测的影响,以淮河王家坝~蒋家集~润河集区间流域为例,设计了七种径流预测方案,采用LSTM(长短期记忆神经网络)、RF(随机森林)以及SVR(支持向量回归)三种机器学习模型进行径流预测。研究结果表明:(1)三种机器学习模型对降雨信息的敏感程度不同,且采用同时考虑径流影响因素以及前期历史径流的方案预测效果最佳,但随着预见期的延长,前期历史径流的重要性逐渐降低;(2)三种机器学习模型在不同预见期的径流预测表现有所差异:三种机器学习模型在预见期为1 d时预测精度均较高;当预见期为2~4 d时,SVR模型的预测效果较好;RF模型在预见期为5~7 d时预测精度较高。研究可为后续基于机器学习的径流预测提供参考。

To explore the impact of different prediction schemes on the runoff prediction of machine learning models,the interval watershed(Wangjiaba-Jiangjiaji-Runheji) of Huaihe River Basin was taken as an example.This study designed seven runoff prediction schemes and used three machine learning models:LSTM(Long Short-Term Memory neural network),RF(Random Forest) and SVR(Support Vector Regression),to predict the runoff in the interval watershed.The results show:(1) The sensitives of three machine learning models to rainfall information are different.The best scheme is that considering both runoff influencing factors and prior historical runoff.However,the importance of prior historical runoff diminishes when extending lead time.(2) The performance of the three machine learning models for runoff prediction is different with different lead time.The three machine learning models all perform well with 1-day lead time,SVR performs better with 2-4 days lead time,and the prediction accuracy of RF is the best when lead time is 5-7 days.The study can provide references for the runoff prediction based on the machine learning.