训练数据量对LSTM网络学习性能影响分析

Effects of Training Data on the Study Performance of LSTM Network

以雅砻江岷江和嘉陵江为研究流域,采用K-最近邻(KNN)算法模拟生成130年的气象数据,并采用SWAT模型计算各流域出口水文站的径流过程;然后分别以前5年、10年、20年、40年和80年的降雨和径流数据对网络进行训练,以最后50年数据作为验证。主要结果表明:LSTM网络的学习能力随着神经元数量增加不断提高,但对水文序列数据的学习则存在过拟合严重的问题;增加训练数据量,可以有效地降低LSTM网络过拟合现象。

This paper took the Yalong,Min and Jialing rivers as the study basins,and applied the K-Nearest Neighbor(KNN)algorithm to simulate and generate 130 years of meteorological data and employed the SWAT model to calculate the runoff process at the outlet hydrological stations of each basin;Then,it trained the network with 5,10,20,40 and 80 years of rainfall and runoff data respectively while used 50 years of data as the validation.The main results show that,(1)The learning ability of the LSTM network continues to improve with the increase in the number of neurons,but there is a serious problem of overfitting in the learning of hydrological sequence data;(2)increasing the amount of training data can effectively reduce the overfitting phenomenon of LSTM network.