基于AR-RNN的多变量水位预测模型研究

Multivariable water level prediction based on AR-RNN model

影响河流水位的因素众多,鉴于传统的单变量水文预测模型无法充分考虑众多因素,提出了一种基于AR-RNN的多变量水位预测模型。模型包含循环神经网络(RNN)与自回归模型(AR)两个部分。RNN部分为模型引入了大量的非线性层,帮助模型拟合水文序列中的非线性成分。但是大量的非线性层降低了模型对于线性成分的敏感性,AR部分可以提高模型对于线性成分的敏感性,使得模型在水位峰值处的预测更加准确。将AR-RNN模型应用于四川省清溪河流域的水位预测中,结果表明:相对于ARIMA模型、SVR模型和BP神经网络,AR-RNN模型的预测精度更高。

There are many factors that affect the water level of a river.In view of the fact that traditional univariate hydrological prediction model cannot fully consider those factors,a multivariable water level prediction model based on AR-RNN was proposed.The model contained two parts,i.e.,RNN and AR.RNN can introduces a large number of non-linear layers to help the model fit the non-linear components in the hydrological sequence.However,more non-linear layers reduces the sensitivity of model to linear components.AR can increase the model sensitivity to linear components,which can make the model's prediction at the peak of the water level to be more accurate.We applied the AR-RNN model to the water level prediction of the Qingxi River in Sichuan Province.The results showed that the AR-RNN model has higher prediction accuracy than the ARIMA model,SVR model and BP neural network.